Daily Papers

Understanding the structure of complex activities in untrimmed videos is a challenging task in the area of action recognition. One problem here is that this task usually requires a large amount of hand-annotated minute- or even hour-long video data, but annotating such data is very time consuming and can not easily be automated or scaled. To address this problem, this paper proposes an approach for the unsupervised learning of actions in untrimmed video sequences based on a joint visual-temporal embedding space. To this end, we combine a visual embedding based on a predictive U-Net architecture with a temporal continuous function. The resulting representation space allows detecting relevant action clusters based on their visual as well as their temporal appearance. The proposed method is evaluated on three standard benchmark datasets, Breakfast Actions, INRIA YouTube Instructional Videos, and 50 Salads. We show that the proposed approach is able to provide a meaningful visual and temporal embedding out of the visual cues present in contiguous video frames and is suitable for the task of unsupervised temporal segmentation of actions.

We present a large-scale study on unsupervised spatiotemporal representation learning from videos. With a unified perspective on four recent image-based frameworks, we study a simple objective that can easily generalize all these methods to space-time. Our objective encourages temporally-persistent features in the same video, and in spite of its simplicity, it works surprisingly well across: (i) different unsupervised frameworks, (ii) pre-training datasets, (iii) downstream datasets, and (iv) backbone architectures. We draw a series of intriguing observations from this study, e.g., we discover that encouraging long-spanned persistency can be effective even if the timespan is 60 seconds. In addition to state-of-the-art results in multiple benchmarks, we report a few promising cases in which unsupervised pre-training can outperform its supervised counterpart. Code is made available at https://github.com/facebookresearch/SlowFast

This paper introduces VLAP, a novel approach that bridges pretrained vision models and large language models (LLMs) to make frozen LLMs understand the visual world. VLAP transforms the embedding space of pretrained vision models into the LLMs' word embedding space using a single linear layer for efficient and general-purpose visual and language understanding. Specifically, we harness well-established word embeddings to bridge two modality embedding spaces. The visual and text representations are simultaneously assigned to a set of word embeddings within pretrained LLMs by formulating the assigning procedure as an optimal transport problem. We predict the assignment of one modality from the representation of another modality data, enforcing consistent assignments for paired multimodal data. This allows vision and language representations to contain the same information, grounding the frozen LLMs' word embedding space in visual data. Moreover, a robust semantic taxonomy of LLMs can be preserved with visual data since the LLMs interpret and reason linguistic information from correlations between word embeddings. Experimental results show that VLAP achieves substantial improvements over the previous linear transformation-based approaches across a range of vision-language tasks, including image captioning, visual question answering, and cross-modal retrieval. We also demonstrate the learned visual representations hold a semantic taxonomy of LLMs, making visual semantic arithmetic possible.

Temporal reasoning is a critical challenge in video-language understanding, as it requires models to align semantic concepts consistently across time. While existing large vision-language models (LVLMs) and large language models (LLMs) excel at static tasks, they struggle to capture dynamic interactions and temporal dependencies in video sequences. In this work, we propose Temporal Semantic Alignment via Dynamic Prompting (TSADP), a novel framework that enhances temporal reasoning capabilities through dynamic task-specific prompts and temporal contrastive learning. TSADP leverages a Dynamic Prompt Generator (DPG) to encode fine-grained temporal relationships and a Temporal Contrastive Loss (TCL) to align visual and textual embeddings across time. We evaluate our method on the VidSitu dataset, augmented with enriched temporal annotations, and demonstrate significant improvements over state-of-the-art models in tasks such as Intra-Video Entity Association, Temporal Relationship Understanding, and Chronology Prediction. Human evaluations further confirm TSADP's ability to generate coherent and semantically accurate descriptions. Our analysis highlights the robustness, efficiency, and practical utility of TSADP, making it a step forward in the field of video-language understanding.

In this paper, we explore the visual representations produced from a pre-trained text-to-video (T2V) diffusion model for video understanding tasks. We hypothesize that the latent representation learned from a pretrained generative T2V model encapsulates rich semantics and coherent temporal correspondences, thereby naturally facilitating video understanding. Our hypothesis is validated through the classic referring video object segmentation (R-VOS) task. We introduce a novel framework, termed "VD-IT", tailored with dedicatedly designed components built upon a fixed pretrained T2V model. Specifically, VD-IT uses textual information as a conditional input, ensuring semantic consistency across time for precise temporal instance matching. It further incorporates image tokens as supplementary textual inputs, enriching the feature set to generate detailed and nuanced masks. Besides, instead of using the standard Gaussian noise, we propose to predict the video-specific noise with an extra noise prediction module, which can help preserve the feature fidelity and elevates segmentation quality. Through extensive experiments, we surprisingly observe that fixed generative T2V diffusion models, unlike commonly used video backbones (e.g., Video Swin Transformer) pretrained with discriminative image/video pre-tasks, exhibit better potential to maintain semantic alignment and temporal consistency. On existing standard benchmarks, our VD-IT achieves highly competitive results, surpassing many existing state-of-the-art methods. The code is available at https://github.com/buxiangzhiren/VD-IT.

Long video understanding remains a fundamental challenge for multimodal large language models (MLLMs), particularly in tasks requiring precise temporal reasoning and event localization. Existing approaches typically adopt uniform frame sampling and rely on implicit position encodings to model temporal order. However, these methods struggle with long-range dependencies, leading to critical information loss and degraded temporal comprehension. In this paper, we propose Dynamic Absolute Time Enhancement (DATE) that enhances temporal awareness in MLLMs through the Timestamp Injection Mechanism (TIM) and a semantically guided Temporal-Aware Similarity Sampling (TASS) strategy. Specifically, we interleave video frame embeddings with textual timestamp tokens to construct a continuous temporal reference system. We further reformulate the video sampling problem as a vision-language retrieval task and introduce a two-stage algorithm to ensure both semantic relevance and temporal coverage: enriching each query into a descriptive caption to better align with the vision feature, and sampling key event with a similarity-driven temporally regularized greedy strategy. Our method achieves remarkable improvements w.r.t. absolute time understanding and key event localization, resulting in state-of-the-art performance among 7B and 72B models on hour-long video benchmarks. Particularly, our 7B model even exceeds many 72B models on some benchmarks.

Pre-training on large-scale video data has become a common recipe for learning transferable spatiotemporal representations in recent years. Despite some progress, existing methods are mostly limited to highly curated datasets (e.g., K400) and exhibit unsatisfactory out-of-the-box representations. We argue that it is due to the fact that they only capture pixel-level knowledge rather than spatiotemporal semantics, which hinders further progress in video understanding. Inspired by the great success of image-text pre-training (e.g., CLIP), we take the first step to exploit language semantics to boost transferable spatiotemporal representation learning. We introduce a new pretext task, Turning to Video for Transcript Sorting (TVTS), which sorts shuffled ASR scripts by attending to learned video representations. We do not rely on descriptive captions and learn purely from video, i.e., leveraging the natural transcribed speech knowledge to provide noisy but useful semantics over time. Our method enforces the vision model to contextualize what is happening over time so that it can re-organize the narrative transcripts, and can seamlessly apply to large-scale uncurated video data in the real world. Our method demonstrates strong out-of-the-box spatiotemporal representations on diverse benchmarks, e.g., +13.6% gains over VideoMAE on SSV2 via linear probing. The code is available at https://github.com/TencentARC/TVTS.

Image-based visual-language (I-VL) pre-training has shown great success for learning joint visual-textual representations from large-scale web data, revealing remarkable ability for zero-shot generalisation. This paper presents a simple but strong baseline to efficiently adapt the pre-trained I-VL model, and exploit its powerful ability for resource-hungry video understanding tasks, with minimal training. Specifically, we propose to optimise a few random vectors, termed as continuous prompt vectors, that convert video-related tasks into the same format as the pre-training objectives. In addition, to bridge the gap between static images and videos, temporal information is encoded with lightweight Transformers stacking on top of frame-wise visual features. Experimentally, we conduct extensive ablation studies to analyse the critical components. On 10 public benchmarks of action recognition, action localisation, and text-video retrieval, across closed-set, few-shot, and zero-shot scenarios, we achieve competitive or state-of-the-art performance to existing methods, despite optimising significantly fewer parameters.

Efficient understanding of long-form videos remains a significant challenge in computer vision. In this work, we revisit temporal search paradigms for long-form video understanding, studying a fundamental issue pertaining to all state-of-the-art (SOTA) long-context vision-language models (VLMs). In particular, our contributions are two-fold: First, we formulate temporal search as a Long Video Haystack problem, i.e., finding a minimal set of relevant frames (typically one to five) among tens of thousands of frames from real-world long videos given specific queries. To validate our formulation, we create LV-Haystack, the first benchmark containing 3,874 human-annotated instances with fine-grained evaluation metrics for assessing keyframe search quality and computational efficiency. Experimental results on LV-Haystack highlight a significant research gap in temporal search capabilities, with SOTA keyframe selection methods achieving only 2.1% temporal F1 score on the LVBench subset. Next, inspired by visual search in images, we re-think temporal searching and propose a lightweight keyframe searching framework, T*, which casts the expensive temporal search as a spatial search problem. T* leverages superior visual localization capabilities typically used in images and introduces an adaptive zooming-in mechanism that operates across both temporal and spatial dimensions. Our extensive experiments show that when integrated with existing methods, T* significantly improves SOTA long-form video understanding performance. Specifically, under an inference budget of 32 frames, T* improves GPT-4o's performance from 50.5% to 53.1% and LLaVA-OneVision-72B's performance from 56.5% to 62.4% on LongVideoBench XL subset. Our PyTorch code, benchmark dataset and models are included in the Supplementary material.

This paper studies referring video object segmentation (RVOS) by boosting video-level visual-linguistic alignment. Recent approaches model the RVOS task as a sequence prediction problem and perform multi-modal interaction as well as segmentation for each frame separately. However, the lack of a global view of video content leads to difficulties in effectively utilizing inter-frame relationships and understanding textual descriptions of object temporal variations. To address this issue, we propose Semantic-assisted Object Cluster (SOC), which aggregates video content and textual guidance for unified temporal modeling and cross-modal alignment. By associating a group of frame-level object embeddings with language tokens, SOC facilitates joint space learning across modalities and time steps. Moreover, we present multi-modal contrastive supervision to help construct well-aligned joint space at the video level. We conduct extensive experiments on popular RVOS benchmarks, and our method outperforms state-of-the-art competitors on all benchmarks by a remarkable margin. Besides, the emphasis on temporal coherence enhances the segmentation stability and adaptability of our method in processing text expressions with temporal variations. Code will be available.

Deep saliency prediction algorithms complement the object recognition features, they typically rely on additional information, such as scene context, semantic relationships, gaze direction, and object dissimilarity. However, none of these models consider the temporal nature of gaze shifts during image observation. We introduce a novel saliency prediction model that learns to output saliency maps in sequential time intervals by exploiting human temporal attention patterns. Our approach locally modulates the saliency predictions by combining the learned temporal maps. Our experiments show that our method outperforms the state-of-the-art models, including a multi-duration saliency model, on the SALICON benchmark. Our code will be publicly available on GitHub.

The correlation between the vision and text is essential for video moment retrieval (VMR), however, existing methods heavily rely on separate pre-training feature extractors for visual and textual understanding. Without sufficient temporal boundary annotations, it is non-trivial to learn universal video-text alignments. In this work, we explore multi-modal correlations derived from large-scale image-text data to facilitate generalisable VMR. To address the limitations of image-text pre-training models on capturing the video changes, we propose a generic method, referred to as Visual-Dynamic Injection (VDI), to empower the model's understanding of video moments. Whilst existing VMR methods are focusing on building temporal-aware video features, being aware of the text descriptions about the temporal changes is also critical but originally overlooked in pre-training by matching static images with sentences. Therefore, we extract visual context and spatial dynamic information from video frames and explicitly enforce their alignments with the phrases describing video changes (e.g. verb). By doing so, the potentially relevant visual and motion patterns in videos are encoded in the corresponding text embeddings (injected) so to enable more accurate video-text alignments. We conduct extensive experiments on two VMR benchmark datasets (Charades-STA and ActivityNet-Captions) and achieve state-of-the-art performances. Especially, VDI yields notable advantages when being tested on the out-of-distribution splits where the testing samples involve novel scenes and vocabulary.

Large video-language models (LVLMs) have shown remarkable performance across various video-language tasks. However, they encounter significant challenges when processing long videos because of the large number of video frames involved. Downsampling long videos in either space or time can lead to visual hallucinations, making it difficult to accurately interpret long videos. Motivated by human hierarchical temporal search strategies, we propose TimeSearch, a novel framework enabling LVLMs to understand long videos in a human-like manner. TimeSearch integrates two human-like primitives into a unified autoregressive LVLM: 1) Spotlight efficiently identifies relevant temporal events through a Temporal-Augmented Frame Representation (TAFR), explicitly binding visual features with timestamps; 2) Reflection evaluates the correctness of the identified events, leveraging the inherent temporal self-reflection capabilities of LVLMs. TimeSearch progressively explores key events and prioritizes temporal search based on reflection confidence. Extensive experiments on challenging long-video benchmarks confirm that TimeSearch substantially surpasses previous state-of-the-art, improving the accuracy from 41.8\% to 51.5\% on the LVBench. Additionally, experiments on temporal grounding demonstrate that appropriate TAFR is adequate to effectively stimulate the surprising temporal grounding ability of LVLMs in a simpler yet versatile manner, which improves mIoU on Charades-STA by 11.8\%. The code will be released.

The alignment of representations from different modalities has recently been shown to provide insights on the structural similarities and downstream capabilities of different encoders across diverse data types. While significant progress has been made in aligning images with text, the temporal nature of video data remains largely unexplored in this context. In this work, we conduct the first comprehensive study of video-text representation alignment, probing the capabilities of modern video and language encoders. Our findings reveal several key insights. First, we demonstrate that cross-modal alignment highly depends on the richness of both visual (static images vs. multi-frame videos) and text (single caption vs. a collection) data provided at test time, especially when using state-of-the-art video encoders. We propose parametric test-time scaling laws that capture this behavior and show remarkable predictive power against empirical observations. Secondly, we investigate the correlation between semantic alignment and performance on both semantic and non-semantic downstream tasks, providing initial evidence that strong alignment against text encoders may be linked to general-purpose video representation and understanding. Finally, we correlate temporal reasoning with cross-modal alignment providing a challenging test-bed for vision and language models. Overall, our work introduces video-text alignment as an informative zero-shot way to probe the representation power of different encoders for spatio-temporal data. Project page can be found at https://video-prh.github.io/

Vision-language-action (VLA) models show potential for general robotic tasks, but remain challenging in spatiotemporally coherent manipulation, which requires fine-grained representations. Typically, existing methods embed 3D positions into visual representations to enhance the spatial precision of actions. However, these methods struggle to achieve temporally coherent control over action execution. In this work, we propose VLA-4D, a general VLA model with 4D awareness for spatiotemporally coherent robotic manipulation. Our model is guided by two key designs: 1) 4D-aware visual representation. We extract visual features, embed 1D time into 3D positions for 4D embeddings, and fuse them into a unified visual representation via a cross-attention mechanism. 2) Spatiotemporal action representation. We extend conventional spatial action representations with temporal information to enable the spatiotemporal planning, and align the multimodal representations into the LLM for spatiotemporal action prediction. Within this unified framework, the designed visual and action representations jointly make robotic manipulation spatially-smooth and temporally-coherent. In addition, we extend the VLA dataset with temporal action annotations for fine-tuning our model. Extensive experiments have been conducted to verify the superiority of our method across different tasks of robotic manipulation.

This paper introduces a novel method for self-supervised video representation learning via feature prediction. In contrast to the previous methods that focus on future feature prediction, we argue that a supervisory signal arising from unobserved past frames is complementary to one that originates from the future frames. The rationale behind our method is to encourage the network to explore the temporal structure of videos by distinguishing between future and past given present observations. We train our model in a contrastive learning framework, where joint encoding of future and past provides us with a comprehensive set of temporal hard negatives via swapping. We empirically show that utilizing both signals enriches the learned representations for the downstream task of action recognition. It outperforms independent prediction of future and past.

Multimodal self-supervised learning is getting more and more attention as it allows not only to train large networks without human supervision but also to search and retrieve data across various modalities. In this context, this paper proposes a self-supervised training framework that learns a common multimodal embedding space that, in addition to sharing representations across different modalities, enforces a grouping of semantically similar instances. To this end, we extend the concept of instance-level contrastive learning with a multimodal clustering step in the training pipeline to capture semantic similarities across modalities. The resulting embedding space enables retrieval of samples across all modalities, even from unseen datasets and different domains. To evaluate our approach, we train our model on the HowTo100M dataset and evaluate its zero-shot retrieval capabilities in two challenging domains, namely text-to-video retrieval, and temporal action localization, showing state-of-the-art results on four different datasets.

Our objective in this work is video-text retrieval - in particular a joint embedding that enables efficient text-to-video retrieval. The challenges in this area include the design of the visual architecture and the nature of the training data, in that the available large scale video-text training datasets, such as HowTo100M, are noisy and hence competitive performance is achieved only at scale through large amounts of compute. We address both these challenges in this paper. We propose an end-to-end trainable model that is designed to take advantage of both large-scale image and video captioning datasets. Our model is an adaptation and extension of the recent ViT and Timesformer architectures, and consists of attention in both space and time. The model is flexible and can be trained on both image and video text datasets, either independently or in conjunction. It is trained with a curriculum learning schedule that begins by treating images as 'frozen' snapshots of video, and then gradually learns to attend to increasing temporal context when trained on video datasets. We also provide a new video-text pretraining dataset WebVid-2M, comprised of over two million videos with weak captions scraped from the internet. Despite training on datasets that are an order of magnitude smaller, we show that this approach yields state-of-the-art results on standard downstream video-retrieval benchmarks including MSR-VTT, MSVD, DiDeMo and LSMDC.

Building joint representations across images and text is an essential step for tasks such as Visual Question Answering and Video Question Answering. In this work, we find that the representations must not only jointly capture features from both modalities but should also be diverse for better generalization performance. To this end, we propose joint vision-language representation learning by diversifying the tokenization learning process, enabling tokens that are sufficiently disentangled from each other to be learned from both modalities. We observe that our approach outperforms the baseline models in a majority of settings and is competitive with state-of-the-art methods.

Models based on deep convolutional networks have dominated recent image interpretation tasks; we investigate whether models which are also recurrent, or "temporally deep", are effective for tasks involving sequences, visual and otherwise. We develop a novel recurrent convolutional architecture suitable for large-scale visual learning which is end-to-end trainable, and demonstrate the value of these models on benchmark video recognition tasks, image description and retrieval problems, and video narration challenges. In contrast to current models which assume a fixed spatio-temporal receptive field or simple temporal averaging for sequential processing, recurrent convolutional models are "doubly deep"' in that they can be compositional in spatial and temporal "layers". Such models may have advantages when target concepts are complex and/or training data are limited. Learning long-term dependencies is possible when nonlinearities are incorporated into the network state updates. Long-term RNN models are appealing in that they directly can map variable-length inputs (e.g., video frames) to variable length outputs (e.g., natural language text) and can model complex temporal dynamics; yet they can be optimized with backpropagation. Our recurrent long-term models are directly connected to modern visual convnet models and can be jointly trained to simultaneously learn temporal dynamics and convolutional perceptual representations. Our results show such models have distinct advantages over state-of-the-art models for recognition or generation which are separately defined and/or optimized.

Current Visual Question Answering (VQA) systems can answer intelligent questions about `Known' visual content. However, their performance drops significantly when questions about visually and linguistically `Unknown' concepts are presented during inference (`Open-world' scenario). A practical VQA system should be able to deal with novel concepts in real world settings. To address this problem, we propose an exemplar-based approach that transfers learning (i.e., knowledge) from previously `Known' concepts to answer questions about the `Unknown'. We learn a highly discriminative joint embedding space, where visual and semantic features are fused to give a unified representation. Once novel concepts are presented to the model, it looks for the closest match from an exemplar set in the joint embedding space. This auxiliary information is used alongside the given Image-Question pair to refine visual attention in a hierarchical fashion. Since handling the high dimensional exemplars on large datasets can be a significant challenge, we introduce an efficient matching scheme that uses a compact feature description for search and retrieval. To evaluate our model, we propose a new split for VQA, separating Unknown visual and semantic concepts from the training set. Our approach shows significant improvements over state-of-the-art VQA models on the proposed Open-World VQA dataset and standard VQA datasets.

We show how perceptual embeddings of the visual system can be constructed at inference-time with no training data or deep neural network features. Our perceptual embeddings are solutions to a weighted least squares (WLS) problem, defined at the pixel-level, and solved at inference-time, that can capture global and local image characteristics. The distance in embedding space is used to define a perceptual similarity metric which we call LASI: Linear Autoregressive Similarity Index. Experiments on full-reference image quality assessment datasets show LASI performs competitively with learned deep feature based methods like LPIPS (Zhang et al., 2018) and PIM (Bhardwaj et al., 2020), at a similar computational cost to hand-crafted methods such as MS-SSIM (Wang et al., 2003). We found that increasing the dimensionality of the embedding space consistently reduces the WLS loss while increasing performance on perceptual tasks, at the cost of increasing the computational complexity. LASI is fully differentiable, scales cubically with the number of embedding dimensions, and can be parallelized at the pixel-level. A Maximum Differentiation (MAD) competition (Wang & Simoncelli, 2008) between LASI and LPIPS shows that both methods are capable of finding failure points for the other, suggesting these metrics can be combined.

Evaluating whether vision-language models (VLMs) reason consistently across representations is challenging because modality comparisons are typically confounded by task differences and asymmetric information. We introduce SEAM, a benchmark that pairs semantically equivalent inputs across four domains that have existing standardized textual and visual notations. By employing distinct notation systems across modalities, in contrast to OCR-based image-text pairing, SEAM provides a rigorous comparative assessment of the textual-symbolic and visual-spatial reasoning capabilities of VLMs. Across 21 contemporary models, we observe systematic modality imbalance: vision frequently lags language in overall performance, despite the problems containing semantically equivalent information, and cross-modal agreement is relatively low. Our error analysis reveals two main drivers: textual perception failures from tokenization in domain notation and visual perception failures that induce hallucinations. We also show that our results are largely robust to visual transformations. SEAM establishes a controlled, semantically equivalent setting for measuring and improving modality-agnostic reasoning.

A key challenge with procedure planning in instructional videos lies in how to handle a large decision space consisting of a multitude of action types that belong to various tasks. To understand real-world video content, an AI agent must proficiently discern these action types (e.g., pour milk, pour water, open lid, close lid, etc.) based on brief visual observation. Moreover, it must adeptly capture the intricate semantic relation of the action types and task goals, along with the variable action sequences. Recently, notable progress has been made via the integration of diffusion models and visual representation learning to address the challenge. However, existing models employ rudimentary mechanisms to utilize task information to manage the decision space. To overcome this limitation, we introduce a simple yet effective enhancement - a masked diffusion model. The introduced mask acts akin to a task-oriented attention filter, enabling the diffusion/denoising process to concentrate on a subset of action types. Furthermore, to bolster the accuracy of task classification, we harness more potent visual representation learning techniques. In particular, we learn a joint visual-text embedding, where a text embedding is generated by prompting a pre-trained vision-language model to focus on human actions. We evaluate the method on three public datasets and achieve state-of-the-art performance on multiple metrics. Code is available at https://github.com/ffzzy840304/Masked-PDPP.

Composed Video Retrieval (CoVR) retrieves a target video given a query video and a modification text describing the intended change. Existing CoVR benchmarks emphasize appearance shifts or coarse event changes and therefore do not test the ability to capture subtle, fast-paced temporal differences. We introduce TF-CoVR, the first large-scale benchmark dedicated to temporally fine-grained CoVR. TF-CoVR focuses on gymnastics and diving and provides 180K triplets drawn from FineGym and FineDiving. Previous CoVR benchmarks focusing on temporal aspect, link each query to a single target segment taken from the same video, limiting practical usefulness. In TF-CoVR, we instead construct each <query, modification> pair by prompting an LLM with the label differences between clips drawn from different videos; every pair is thus associated with multiple valid target videos (3.9 on average), reflecting real-world tasks such as sports-highlight generation. To model these temporal dynamics we propose TF-CoVR-Base, a concise two-stage training framework: (i) pre-train a video encoder on fine-grained action classification to obtain temporally discriminative embeddings; (ii) align the composed query with candidate videos using contrastive learning. We conduct the first comprehensive study of image, video, and general multimodal embedding (GME) models on temporally fine-grained composed retrieval in both zero-shot and fine-tuning regimes. On TF-CoVR, TF-CoVR-Base improves zero-shot mAP@50 from 5.92 (LanguageBind) to 7.51, and after fine-tuning raises the state-of-the-art from 19.83 to 25.82.

There has been tremendous progress in multimodal Large Language Models (LLMs). Recent works have extended these models to video input with promising instruction following capabilities. However, an important missing piece is temporal localization. These models cannot accurately answer the "When?" questions. We identify three key aspects that limit their temporal localization capabilities: (i) time representation, (ii) architecture, and (iii) data. We address these shortcomings by proposing Language Instructed Temporal-Localization Assistant (LITA) with the following features: (1) We introduce time tokens that encode timestamps relative to the video length to better represent time in videos. (2) We introduce SlowFast tokens in the architecture to capture temporal information at fine temporal resolution. (3) We emphasize temporal localization data for LITA. In addition to leveraging existing video datasets with timestamps, we propose a new task, Reasoning Temporal Localization (RTL), along with the dataset, ActivityNet-RTL, for learning and evaluating this task. Reasoning temporal localization requires both the reasoning and temporal localization of Video LLMs. LITA demonstrates strong performance on this challenging task, nearly doubling the temporal mean intersection-over-union (mIoU) of baselines. In addition, we show that our emphasis on temporal localization also substantially improves video-based text generation compared to existing Video LLMs, including a 36% relative improvement of Temporal Understanding. Code is available at: https://github.com/NVlabs/LITA

Understanding videos requires more than answering open ended questions, it demands the ability to pinpoint when events occur and how entities interact across time. While recent Video LLMs have achieved remarkable progress in holistic reasoning, they remain coarse in temporal perception: timestamps are encoded only implicitly, frame level features are weak in capturing continuity, and language vision alignment often drifts from the entities of interest. In this paper, we present Grounded VideoDiT, a Video LLM designed to overcome these limitations by introducing three key innovations. First, a Diffusion Temporal Latent (DTL) encoder enhances boundary sensitivity and maintains temporal consistency. Second, object grounded representations explicitly bind query entities to localized visual evidence, strengthening alignment. Third, a mixed token scheme with discrete temporal tokens provides explicit timestamp modeling, enabling fine grained temporal reasoning. Together, these designs equip Grounded VideoDiT with robust grounding capabilities, as validated by state of the art results on Charades STA, NExT GQA, and multiple VideoQA benchmarks.

Multimodal embedding models have been crucial in enabling various downstream tasks such as semantic similarity, information retrieval, and clustering over different modalities. However, existing multimodal embeddings like VLM2Vec, E5-V, GME are predominantly focused on natural images, with limited support for other visual forms such as videos and visual documents. This restricts their applicability in real-world scenarios, including AI agents, multi-modal search and recommendation, and retrieval-augmented generation (RAG). To close this gap, we propose VLM2Vec-V2, a unified framework for learning embeddings across diverse visual forms. First, we introduce MMEB-V2, a comprehensive benchmark that extends MMEB with five new task types: visual document retrieval, video retrieval, temporal grounding, video classification and video question answering - spanning text, image, video, and visual document inputs. Next, we train VLM2Vec-V2, a general-purpose embedding model that supports text, image, video, and visual document inputs. Extensive experiments show that VLM2Vec-V2 achieves strong performance not only on the newly introduced video and document retrieval tasks, but also improves over prior baselines on the original image benchmarks. Through extensive evaluation, our study offers insights into the generalizability of various multimodal embedding models and highlights effective strategies for unified embedding learning, laying the groundwork for more scalable and adaptable representation learning in both research and real-world settings.

Self-supervised methods have shown remarkable progress in learning high-level semantics and low-level temporal correspondence. Building on these results, we take one step further and explore the possibility of integrating these two features to enhance object-centric representations. Our preliminary experiments indicate that query slot attention can extract different semantic components from the RGB feature map, while random sampling based slot attention can exploit temporal correspondence cues between frames to assist instance identification. Motivated by this, we propose a novel semantic-aware masked slot attention on top of the fused semantic features and correspondence maps. It comprises two slot attention stages with a set of shared learnable Gaussian distributions. In the first stage, we use the mean vectors as slot initialization to decompose potential semantics and generate semantic segmentation masks through iterative attention. In the second stage, for each semantics, we randomly sample slots from the corresponding Gaussian distribution and perform masked feature aggregation within the semantic area to exploit temporal correspondence patterns for instance identification. We adopt semantic- and instance-level temporal consistency as self-supervision to encourage temporally coherent object-centric representations. Our model effectively identifies multiple object instances with semantic structure, reaching promising results on unsupervised video object discovery. Furthermore, we achieve state-of-the-art performance on dense label propagation tasks, demonstrating the potential for object-centric analysis. The code is released at https://github.com/shvdiwnkozbw/SMTC.

We present the task of Spatio-Temporal Video Question Answering, which requires intelligent systems to simultaneously retrieve relevant moments and detect referenced visual concepts (people and objects) to answer natural language questions about videos. We first augment the TVQA dataset with 310.8K bounding boxes, linking depicted objects to visual concepts in questions and answers. We name this augmented version as TVQA+. We then propose Spatio-Temporal Answerer with Grounded Evidence (STAGE), a unified framework that grounds evidence in both spatial and temporal domains to answer questions about videos. Comprehensive experiments and analyses demonstrate the effectiveness of our framework and how the rich annotations in our TVQA+ dataset can contribute to the question answering task. Moreover, by performing this joint task, our model is able to produce insightful and interpretable spatio-temporal attention visualizations. Dataset and code are publicly available at: http: //tvqa.cs.unc.edu, https://github.com/jayleicn/TVQAplus

Several video understanding tasks, such as natural language temporal video grounding, temporal activity localization, and audio description generation, require "temporally dense" reasoning over frames sampled at high temporal resolution. However, computing frame-level features for these tasks is computationally expensive given the temporal resolution requirements. In this paper, we make three contributions to reduce the cost of computing features for temporally dense tasks. First, we introduce a vision transformer (ViT) architecture, dubbed ResidualViT, that leverages the large temporal redundancy in videos to efficiently compute temporally dense frame-level features. Our architecture incorporates (i) learnable residual connections that ensure temporal consistency across consecutive frames and (ii) a token reduction module that enhances processing speed by selectively discarding temporally redundant information while reusing weights of a pretrained foundation model. Second, we propose a lightweight distillation strategy to approximate the frame-level features of the original foundation model. Finally, we evaluate our approach across four tasks and five datasets, in both zero-shot and fully supervised settings, demonstrating significant reductions in computational cost (up to 60%) and improvements in inference speed (up to 2.5x faster), all while closely approximating the accuracy of the original foundation model.

Audio-visual generalised zero-shot learning for video classification requires understanding the relations between the audio and visual information in order to be able to recognise samples from novel, previously unseen classes at test time. The natural semantic and temporal alignment between audio and visual data in video data can be exploited to learn powerful representations that generalise to unseen classes at test time. We propose a multi-modal and Temporal Cross-attention Framework (\modelName) for audio-visual generalised zero-shot learning. Its inputs are temporally aligned audio and visual features that are obtained from pre-trained networks. Encouraging the framework to focus on cross-modal correspondence across time instead of self-attention within the modalities boosts the performance significantly. We show that our proposed framework that ingests temporal features yields state-of-the-art performance on the \ucf, \vgg, and \activity benchmarks for (generalised) zero-shot learning. Code for reproducing all results is available at https://github.com/ExplainableML/TCAF-GZSL.

Recent progress in using recurrent neural networks (RNNs) for image description has motivated the exploration of their application for video description. However, while images are static, working with videos requires modeling their dynamic temporal structure and then properly integrating that information into a natural language description. In this context, we propose an approach that successfully takes into account both the local and global temporal structure of videos to produce descriptions. First, our approach incorporates a spatial temporal 3-D convolutional neural network (3-D CNN) representation of the short temporal dynamics. The 3-D CNN representation is trained on video action recognition tasks, so as to produce a representation that is tuned to human motion and behavior. Second we propose a temporal attention mechanism that allows to go beyond local temporal modeling and learns to automatically select the most relevant temporal segments given the text-generating RNN. Our approach exceeds the current state-of-art for both BLEU and METEOR metrics on the Youtube2Text dataset. We also present results on a new, larger and more challenging dataset of paired video and natural language descriptions.

Recent two-stream deep Convolutional Neural Networks (ConvNets) have made significant progress in recognizing human actions in videos. Despite their success, methods extending the basic two-stream ConvNet have not systematically explored possible network architectures to further exploit spatiotemporal dynamics within video sequences. Further, such networks often use different baseline two-stream networks. Therefore, the differences and the distinguishing factors between various methods using Recurrent Neural Networks (RNN) or convolutional networks on temporally-constructed feature vectors (Temporal-ConvNet) are unclear. In this work, we first demonstrate a strong baseline two-stream ConvNet using ResNet-101. We use this baseline to thoroughly examine the use of both RNNs and Temporal-ConvNets for extracting spatiotemporal information. Building upon our experimental results, we then propose and investigate two different networks to further integrate spatiotemporal information: 1) temporal segment RNN and 2) Inception-style Temporal-ConvNet. We demonstrate that using both RNNs (using LSTMs) and Temporal-ConvNets on spatiotemporal feature matrices are able to exploit spatiotemporal dynamics to improve the overall performance. However, each of these methods require proper care to achieve state-of-the-art performance; for example, LSTMs require pre-segmented data or else they cannot fully exploit temporal information. Our analysis identifies specific limitations for each method that could form the basis of future work. Our experimental results on UCF101 and HMDB51 datasets achieve state-of-the-art performances, 94.1% and 69.0%, respectively, without requiring extensive temporal augmentation.

We present VideoFactory, an innovative framework for generating high-quality open-domain videos. VideoFactory excels in producing high-definition (1376x768), widescreen (16:9) videos without watermarks, creating an engaging user experience. Generating videos guided by text instructions poses significant challenges, such as modeling the complex relationship between space and time, and the lack of large-scale text-video paired data. Previous approaches extend pretrained text-to-image generation models by adding temporal 1D convolution/attention modules for video generation. However, these approaches overlook the importance of jointly modeling space and time, inevitably leading to temporal distortions and misalignment between texts and videos. In this paper, we propose a novel approach that strengthens the interaction between spatial and temporal perceptions. In particular, we utilize a swapped cross-attention mechanism in 3D windows that alternates the "query" role between spatial and temporal blocks, enabling mutual reinforcement for each other. To fully unlock model capabilities for high-quality video generation, we curate a large-scale video dataset called HD-VG-130M. This dataset comprises 130 million text-video pairs from the open-domain, ensuring high-definition, widescreen and watermark-free characters. Objective metrics and user studies demonstrate the superiority of our approach in terms of per-frame quality, temporal correlation, and text-video alignment, with clear margins.

Applying Multimodal Large Language Models (MLLMs) to video understanding presents significant challenges due to the need to model temporal relations across frames. Existing approaches adopt either implicit temporal modeling, relying solely on the LLM decoder, or explicit temporal modeling, employing auxiliary temporal encoders. To investigate this debate between the two paradigms, we propose the Stackable Temporal Encoder (STE). STE enables flexible explicit temporal modeling with adjustable temporal receptive fields and token compression ratios. Using STE, we systematically compare implicit and explicit temporal modeling across dimensions such as overall performance, token compression effectiveness, and temporal-specific understanding. We also explore STE's design considerations and broader impacts as a plug-in module and in image modalities. Our findings emphasize the critical role of explicit temporal modeling, providing actionable insights to advance video MLLMs.

Video-text retrieval is a class of cross-modal representation learning problems, where the goal is to select the video which corresponds to the text query between a given text query and a pool of candidate videos. The contrastive paradigm of vision-language pretraining has shown promising success with large-scale datasets and unified transformer architecture, and demonstrated the power of a joint latent space. Despite this, the intrinsic divergence between the visual domain and textual domain is still far from being eliminated, and projecting different modalities into a joint latent space might result in the distorting of the information inside the single modality. To overcome the above issue, we present a novel mechanism for learning the translation relationship from a source modality space S to a target modality space T without the need for a joint latent space, which bridges the gap between visual and textual domains. Furthermore, to keep cycle consistency between translations, we adopt a cycle loss involving both forward translations from S to the predicted target space T', and backward translations from T' back to S. Extensive experiments conducted on MSR-VTT, MSVD, and DiDeMo datasets demonstrate the superiority and effectiveness of our LaT approach compared with vanilla state-of-the-art methods.

In this research, we present a novel approach to motion customization in video generation, addressing the widespread gap in the thorough exploration of motion representation within video generative models. Recognizing the unique challenges posed by video's spatiotemporal nature, our method introduces Motion Embeddings, a set of explicit, temporally coherent one-dimensional embeddings derived from a given video. These embeddings are designed to integrate seamlessly with the temporal transformer modules of video diffusion models, modulating self-attention computations across frames without compromising spatial integrity. Our approach offers a compact and efficient solution to motion representation and enables complex manipulations of motion characteristics through vector arithmetic in the embedding space. Furthermore, we identify the Temporal Discrepancy in video generative models, which refers to variations in how different motion modules process temporal relationships between frames. We leverage this understanding to optimize the integration of our motion embeddings. Our contributions include the introduction of a tailored motion embedding for customization tasks, insights into the temporal processing differences in video models, and a demonstration of the practical advantages and effectiveness of our method through extensive experiments.

Egocentric video-language understanding demands both high efficiency and accurate spatial-temporal modeling. Existing approaches face three key challenges: 1) Excessive pre-training cost arising from multi-stage pre-training pipelines, 2) Ineffective spatial-temporal encoding due to manually split 3D rotary positional embeddings that hinder feature interactions, and 3) Imprecise learning objectives in soft-label multi-instance retrieval, which neglect negative pair correlations. In this paper, we introduce EVA02-AT, a suite of EVA02-based video-language foundation models tailored to egocentric video understanding tasks. EVA02-AT first efficiently transfers an image-based CLIP model into a unified video encoder via a single-stage pretraining. Second, instead of applying rotary positional embeddings to isolated dimensions, we introduce spatial-temporal rotary positional embeddings along with joint attention, which can effectively encode both spatial and temporal information on the entire hidden dimension. This joint encoding of spatial-temporal features enables the model to learn cross-axis relationships, which are crucial for accurately modeling motion and interaction in videos. Third, focusing on multi-instance video-language retrieval tasks, we introduce the Symmetric Multi-Similarity (SMS) loss and a novel training framework that advances all soft labels for both positive and negative pairs, providing a more precise learning objective. Extensive experiments on Ego4D, EPIC-Kitchens-100, and Charades-Ego under zero-shot and fine-tuning settings demonstrate that EVA02-AT achieves state-of-the-art performance across diverse egocentric video-language tasks with fewer parameters. Models with our SMS loss also show significant performance gains on multi-instance retrieval benchmarks. Our code and models are publicly available at https://github.com/xqwang14/EVA02-AT .

In this work, we identify an inherent bias in prevailing LVLM architectures toward the language modality, largely resulting from the common practice of simply appending visual embeddings to the input text sequence. To address this, we propose a simple yet effective method that refines textual embeddings by integrating average-pooled visual features. Our approach demonstrably improves visual grounding and significantly reduces hallucinations on established benchmarks. While average pooling offers a straightforward, robust, and efficient means of incorporating visual information, we believe that more sophisticated fusion methods could further enhance visual grounding and cross-modal alignment. Given that the primary focus of this work is to highlight the modality imbalance and its impact on hallucinations -- and to show that refining textual embeddings with visual information mitigates this issue -- we leave exploration of advanced fusion strategies for future work.

We introduce VideoComp, a benchmark and learning framework for advancing video-text compositionality understanding, aimed at improving vision-language models (VLMs) in fine-grained temporal alignment. Unlike existing benchmarks focused on static image-text compositionality or isolated single-event videos, our benchmark targets alignment in continuous multi-event videos. Leveraging video-text datasets with temporally localized event captions (e.g. ActivityNet-Captions, YouCook2), we construct two compositional benchmarks, ActivityNet-Comp and YouCook2-Comp. We create challenging negative samples with subtle temporal disruptions such as reordering, action word replacement, partial captioning, and combined disruptions. These benchmarks comprehensively test models' compositional sensitivity across extended, cohesive video-text sequences. To improve model performance, we propose a hierarchical pairwise preference loss that strengthens alignment with temporally accurate pairs and gradually penalizes increasingly disrupted ones, encouraging fine-grained compositional learning. To mitigate the limited availability of densely annotated video data, we introduce a pretraining strategy that concatenates short video-caption pairs to simulate multi-event sequences. We evaluate video-text foundational models and large multimodal models (LMMs) on our benchmark, identifying both strengths and areas for improvement in compositionality. Overall, our work provides a comprehensive framework for evaluating and enhancing model capabilities in achieving fine-grained, temporally coherent video-text alignment.

Video understanding represents the most challenging frontier in computer vision, requiring models to reason about complex spatiotemporal relationships, long-term dependencies, and multimodal evidence. The recent emergence of Video-Large Multimodal Models (Video-LMMs), which integrate visual encoders with powerful decoder-based language models, has demonstrated remarkable capabilities in video understanding tasks. However, the critical phase that transforms these models from basic perception systems into sophisticated reasoning engines, post-training, remains fragmented across the literature. This survey provides the first comprehensive examination of post-training methodologies for Video-LMMs, encompassing three fundamental pillars: supervised fine-tuning (SFT) with chain-of-thought, reinforcement learning (RL) from verifiable objectives, and test-time scaling (TTS) through enhanced inference computation. We present a structured taxonomy that clarifies the roles, interconnections, and video-specific adaptations of these techniques, addressing unique challenges such as temporal localization, spatiotemporal grounding, long video efficiency, and multimodal evidence integration. Through systematic analysis of representative methods, we synthesize key design principles, insights, and evaluation protocols while identifying critical open challenges in reward design, scalability, and cost-performance optimization. We further curate essential benchmarks, datasets, and metrics to facilitate rigorous assessment of post-training effectiveness. This survey aims to provide researchers and practitioners with a unified framework for advancing Video-LMM capabilities. Additional resources and updates are maintained at: https://github.com/yunlong10/Awesome-Video-LMM-Post-Training

We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, "visual sentences", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion tokens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time.

Recent advances in video-based multimodal large language models (Video-LLMs) have significantly improved video understanding by processing videos as sequences of image frames. However, many existing methods treat frames independently in the vision backbone, lacking explicit temporal modeling, which limits their ability to capture dynamic patterns and efficiently handle long videos. To address these limitations, we introduce STORM (Spatiotemporal TOken Reduction for Multimodal LLMs), a novel architecture incorporating a dedicated temporal encoder between the image encoder and the LLM. Our temporal encoder leverages the Mamba State Space Model to integrate temporal information into image tokens, generating enriched representations that preserve inter-frame dynamics across the entire video sequence. This enriched encoding not only enhances video reasoning capabilities but also enables effective token reduction strategies, including test-time sampling and training-based temporal and spatial pooling, substantially reducing computational demands on the LLM without sacrificing key temporal information. By integrating these techniques, our approach simultaneously reduces training and inference latency while improving performance, enabling efficient and robust video understanding over extended temporal contexts. Extensive evaluations show that STORM achieves state-of-the-art results across various long video understanding benchmarks (more than 5\% improvement on MLVU and LongVideoBench) while reducing the computation costs by up to 8times and the decoding latency by 2.4-2.9times for the fixed numbers of input frames. Project page is available at https://research.nvidia.com/labs/lpr/storm

Transformers have recently been popular for learning and inference in the spatial-temporal domain. However, their performance relies on storing and applying attention to the feature tensor of each frame in video. Hence, their space and time complexity increase linearly as the length of video grows, which could be very costly for long videos. We propose a novel visual memory network architecture for the learning and inference problem in the spatial-temporal domain. We maintain a fixed set of memory slots in our memory network and propose an algorithm based on Gumbel-Softmax to learn an adaptive strategy to update this memory. Finally, this architecture is benchmarked on the video object segmentation (VOS) and video prediction problems. We demonstrate that our memory architecture achieves state-of-the-art results, outperforming transformer-based methods on VOS and other recent methods on video prediction while maintaining constant memory capacity independent of the sequence length.

We present an unsupervised representation learning approach that compactly encodes the motion dependencies in videos. Given a pair of images from a video clip, our framework learns to predict the long-term 3D motions. To reduce the complexity of the learning framework, we propose to describe the motion as a sequence of atomic 3D flows computed with RGB-D modality. We use a Recurrent Neural Network based Encoder-Decoder framework to predict these sequences of flows. We argue that in order for the decoder to reconstruct these sequences, the encoder must learn a robust video representation that captures long-term motion dependencies and spatial-temporal relations. We demonstrate the effectiveness of our learned temporal representations on activity classification across multiple modalities and datasets such as NTU RGB+D and MSR Daily Activity 3D. Our framework is generic to any input modality, i.e., RGB, Depth, and RGB-D videos.

Contrastive language-image pretraining (CLIP) has significantly advanced image-based vision learning. A pressing topic subsequently arises: how can we effectively adapt CLIP to the video domain? Recent studies have focused on adjusting either the textual or visual branch of CLIP for action recognition. However, we argue that adaptations of both branches are crucial. In this paper, we propose CLAVER: a Contrastive Language-Action Video Learner, designed to shift CLIP's focus from the alignment of static visual objects and concrete nouns to the alignment of dynamic action behaviors and abstract verbs. Specifically, we introduce a novel Kronecker mask attention for temporal modeling. Our tailored Kronecker mask offers three benefits 1) it expands the temporal receptive field for each token, 2) it serves as an effective spatiotemporal heterogeneity inductive bias, mitigating the issue of spatiotemporal homogenization, and 3) it can be seamlessly plugged into transformer-based models. Regarding the textual branch, we leverage large language models to generate diverse, sentence-level and semantically rich interpretive prompts of actions, which shift the model's focus towards the verb comprehension. Extensive experiments on various benchmarks and learning scenarios demonstrate the superiority and generality of our approach.

Time series classification is a fundamental task in healthcare and industry, yet the development of time series foundation models (TSFMs) remains limited by the scarcity of publicly available time series datasets. In this work, we propose Time Vision Transformer (TiViT), a framework that converts time series into images to leverage the representational power of frozen Vision Transformers (ViTs) pretrained on large-scale image datasets. First, we theoretically motivate our approach by analyzing the 2D patching of ViTs for time series, showing that it can increase the number of label-relevant tokens and reduce the sample complexity. Second, we empirically demonstrate that TiViT achieves state-of-the-art performance on standard time series classification benchmarks by utilizing the hidden representations of large OpenCLIP models. We explore the structure of TiViT representations and find that intermediate layers with high intrinsic dimension are the most effective for time series classification. Finally, we assess the alignment between TiViT and TSFM representation spaces and identify a strong complementarity, with further performance gains achieved by combining their features. Our findings reveal a new direction for reusing vision representations in a non-visual domain. Code is available at https://github.com/ExplainableML/TiViT.

We build upon time-series classification by leveraging the capabilities of Vision Language Models (VLMs). We find that VLMs produce competitive results after two or less epochs of fine-tuning. We develop a novel approach that incorporates graphical data representations as images in conjunction with numerical data. This approach is rooted in the hypothesis that graphical representations can provide additional contextual information that numerical data alone may not capture. Additionally, providing a graphical representation can circumvent issues such as limited context length faced by LLMs. To further advance this work, we implemented a scalable end-to-end pipeline for training on different scenarios, allowing us to isolate the most effective strategies for transferring learning capabilities from LLMs to Time Series Classification (TSC) tasks. Our approach works with univariate and multivariate time-series data. In addition, we conduct extensive and practical experiments to show how this approach works for time-series classification and generative labels.

Although contrastive and other representation-learning methods have long been explored in vision and NLP, their adoption in modern time series forecasters remains limited. We believe they hold strong promise for this domain. To unlock this potential, we explicitly align past and future representations, thereby bridging the distributional gap between input histories and future targets. To this end, we introduce TimeAlign, a lightweight, plug-and-play framework that establishes a new representation paradigm, distinct from contrastive learning, by aligning auxiliary features via a simple reconstruction task and feeding them back into any base forecaster. Extensive experiments across eight benchmarks verify its superior performance. Further studies indicate that the gains arise primarily from correcting frequency mismatches between historical inputs and future outputs. Additionally, we provide two theoretical justifications for how reconstruction improves forecasting generalization and how alignment increases the mutual information between learned representations and predicted targets. The code is available at https://github.com/TROUBADOUR000/TimeAlign.

We present GDFusion, a temporal fusion method for vision-based 3D semantic occupancy prediction (VisionOcc). GDFusion opens up the underexplored aspects of temporal fusion within the VisionOcc framework, focusing on both temporal cues and fusion strategies. It systematically examines the entire VisionOcc pipeline, identifying three fundamental yet previously overlooked temporal cues: scene-level consistency, motion calibration, and geometric complementation. These cues capture diverse facets of temporal evolution and make distinct contributions across various modules in the VisionOcc framework. To effectively fuse temporal signals across heterogeneous representations, we propose a novel fusion strategy by reinterpreting the formulation of vanilla RNNs. This reinterpretation leverages gradient descent on features to unify the integration of diverse temporal information, seamlessly embedding the proposed temporal cues into the network. Extensive experiments on nuScenes demonstrate that GDFusion significantly outperforms established baselines. Notably, on Occ3D benchmark, it achieves 1.4\%-4.8\% mIoU improvements and reduces memory consumption by 27\%-72\%.

Self-supervised learning in vision-language processing exploits semantic alignment between imaging and text modalities. Prior work in biomedical VLP has mostly relied on the alignment of single image and report pairs even though clinical notes commonly refer to prior images. This does not only introduce poor alignment between the modalities but also a missed opportunity to exploit rich self-supervision through existing temporal content in the data. In this work, we explicitly account for prior images and reports when available during both training and fine-tuning. Our approach, named BioViL-T, uses a CNN-Transformer hybrid multi-image encoder trained jointly with a text model. It is designed to be versatile to arising challenges such as pose variations and missing input images across time. The resulting model excels on downstream tasks both in single- and multi-image setups, achieving state-of-the-art performance on (I) progression classification, (II) phrase grounding, and (III) report generation, whilst offering consistent improvements on disease classification and sentence-similarity tasks. We release a novel multi-modal temporal benchmark dataset, MS-CXR-T, to quantify the quality of vision-language representations in terms of temporal semantics. Our experimental results show the advantages of incorporating prior images and reports to make most use of the data.

With the release of large-scale motion datasets with textual annotations, the task of establishing a robust latent space for language and 3D human motion has recently witnessed a surge of interest. Methods have been proposed to convert human motion and texts into features to achieve accurate correspondence between them. Despite these efforts to align language and motion representations, we claim that the temporal element is often overlooked, especially for compound actions, resulting in chronological inaccuracies. To shed light on the temporal alignment in motion-language latent spaces, we propose Chronologically Accurate Retrieval (CAR) to evaluate the chronological understanding of the models. We decompose textual descriptions into events, and prepare negative text samples by shuffling the order of events in compound action descriptions. We then design a simple task for motion-language models to retrieve the more likely text from the ground truth and its chronologically shuffled version. CAR reveals many cases where current motion-language models fail to distinguish the event chronology of human motion, despite their impressive performance in terms of conventional evaluation metrics. To achieve better temporal alignment between text and motion, we further propose to use these texts with shuffled sequence of events as negative samples during training to reinforce the motion-language models. We conduct experiments on text-motion retrieval and text-to-motion generation using the reinforced motion-language models, which demonstrate improved performance over conventional approaches, indicating the necessity to consider temporal elements in motion-language alignment.

The recent advancement in video temporal grounding (VTG) has significantly enhanced fine-grained video understanding, primarily driven by multimodal large language models (MLLMs). With superior multimodal comprehension and reasoning abilities, VTG approaches based on MLLMs (VTG-MLLMs) are gradually surpassing traditional fine-tuned methods. They not only achieve competitive performance but also excel in generalization across zero-shot, multi-task, and multi-domain settings. Despite extensive surveys on general video-language understanding, comprehensive reviews specifically addressing VTG-MLLMs remain scarce. To fill this gap, this survey systematically examines current research on VTG-MLLMs through a three-dimensional taxonomy: 1) the functional roles of MLLMs, highlighting their architectural significance; 2) training paradigms, analyzing strategies for temporal reasoning and task adaptation; and 3) video feature processing techniques, which determine spatiotemporal representation effectiveness. We further discuss benchmark datasets, evaluation protocols, and summarize empirical findings. Finally, we identify existing limitations and propose promising research directions. For additional resources and details, readers are encouraged to visit our repository at https://github.com/ki-lw/Awesome-MLLMs-for-Video-Temporal-Grounding.

Referring Video Object Segmentation (RefVOS) seeks to segment target objects in videos guided by natural language descriptions, demanding both temporal reasoning and fine-grained visual comprehension. Existing sampling strategies for LLM-based approaches typically rely on either handcrafted heuristics or external keyframe models. The former often overlooks essential temporal cues, while the latter increases system complexity. To address this, we propose a unified framework that jointly optimizes Temporal Sentence Grounding (TSG) and RefVOS, naturally incorporating key moment grounding capability. During training, we introduce a novel TSG paradigm that employs a dedicated [FIND] token for key moment identification through temporal token similarity matching, thereby avoiding the need for external timestamp encodings. For inference, we design a Moment-Centric Sampling (MCS) strategy that densely samples informative moments while sparsely sampling non-essential frames, preserving both motion details and global context. To further enhance tracking stability, we develop Bidirectional Anchor-updated Propagation (BAP), which leverages the most relevant moment as start point for high-quality mask initialization and dynamically updates at sampled points to mitigate accumulated errors. Code and model will be available at: https://github.com/Dmmm1997/MomentSeg

Contrastive language-image pretraining (CLIP) has demonstrated remarkable success in various image tasks. However, how to extend CLIP with effective temporal modeling is still an open and crucial problem. Existing factorized or joint spatial-temporal modeling trades off between the efficiency and performance. While modeling temporal information within straight through tube is widely adopted in literature, we find that simple frame alignment already provides enough essence without temporal attention. To this end, in this paper, we proposed a novel Implicit Learnable Alignment (ILA) method, which minimizes the temporal modeling effort while achieving incredibly high performance. Specifically, for a frame pair, an interactive point is predicted in each frame, serving as a mutual information rich region. By enhancing the features around the interactive point, two frames are implicitly aligned. The aligned features are then pooled into a single token, which is leveraged in the subsequent spatial self-attention. Our method allows eliminating the costly or insufficient temporal self-attention in video. Extensive experiments on benchmarks demonstrate the superiority and generality of our module. Particularly, the proposed ILA achieves a top-1 accuracy of 88.7% on Kinetics-400 with much fewer FLOPs compared with Swin-L and ViViT-H. Code is released at https://github.com/Francis-Rings/ILA .

Recently, there has been a growing interest in leveraging pre-trained large language models (LLMs) for various time series applications. However, the semantic space of LLMs, established through the pre-training, is still underexplored and may help yield more distinctive and informative representations to facilitate time series forecasting. To this end, we propose Semantic Space Informed Prompt learning with LLM (S^2IP-LLM) to align the pre-trained semantic space with time series embeddings space and perform time series forecasting based on learned prompts from the joint space. We first design a tokenization module tailored for cross-modality alignment, which explicitly concatenates patches of decomposed time series components to create embeddings that effectively encode the temporal dynamics. Next, we leverage the pre-trained word token embeddings to derive semantic anchors and align selected anchors with time series embeddings by maximizing the cosine similarity in the joint space. This way, S^2IP-LLM can retrieve relevant semantic anchors as prompts to provide strong indicators (context) for time series that exhibit different temporal dynamics. With thorough empirical studies on multiple benchmark datasets, we demonstrate that the proposed S^2IP-LLM can achieve superior forecasting performance over state-of-the-art baselines. Furthermore, our ablation studies and visualizations verify the necessity of prompt learning informed by semantic space.

We present a self-supervised Contrastive Video Representation Learning (CVRL) method to learn spatiotemporal visual representations from unlabeled videos. Our representations are learned using a contrastive loss, where two augmented clips from the same short video are pulled together in the embedding space, while clips from different videos are pushed away. We study what makes for good data augmentations for video self-supervised learning and find that both spatial and temporal information are crucial. We carefully design data augmentations involving spatial and temporal cues. Concretely, we propose a temporally consistent spatial augmentation method to impose strong spatial augmentations on each frame of the video while maintaining the temporal consistency across frames. We also propose a sampling-based temporal augmentation method to avoid overly enforcing invariance on clips that are distant in time. On Kinetics-600, a linear classifier trained on the representations learned by CVRL achieves 70.4% top-1 accuracy with a 3D-ResNet-50 (R3D-50) backbone, outperforming ImageNet supervised pre-training by 15.7% and SimCLR unsupervised pre-training by 18.8% using the same inflated R3D-50. The performance of CVRL can be further improved to 72.9% with a larger R3D-152 (2x filters) backbone, significantly closing the gap between unsupervised and supervised video representation learning. Our code and models will be available at https://github.com/tensorflow/models/tree/master/official/.

Recent advancements in multimodal large language models (MLLMs) have shown promising results, yet existing approaches struggle to effectively handle both temporal and spatial localization simultaneously. This challenge stems from two key issues: first, incorporating spatial-temporal localization introduces a vast number of coordinate combinations, complicating the alignment of linguistic and visual coordinate representations; second, encoding fine-grained temporal and spatial information during video feature compression is inherently difficult. To address these issues, we propose LLaVA-ST, a MLLM for fine-grained spatial-temporal multimodal understanding. In LLaVA-ST, we propose Language-Aligned Positional Embedding, which embeds the textual coordinate special token into the visual space, simplifying the alignment of fine-grained spatial-temporal correspondences. Additionally, we design the Spatial-Temporal Packer, which decouples the feature compression of temporal and spatial resolutions into two distinct point-to-region attention processing streams. Furthermore, we propose ST-Align dataset with 4.3M training samples for fine-grained spatial-temporal multimodal understanding. With ST-align, we present a progressive training pipeline that aligns the visual and textual feature through sequential coarse-to-fine stages.Additionally, we introduce an ST-Align benchmark to evaluate spatial-temporal interleaved fine-grained understanding tasks, which include Spatial-Temporal Video Grounding (STVG) , Event Localization and Captioning (ELC) and Spatial Video Grounding (SVG). LLaVA-ST achieves outstanding performance on 11 benchmarks requiring fine-grained temporal, spatial, or spatial-temporal interleaving multimodal understanding. Our code, data and benchmark will be released at Our code, data and benchmark will be released at https://github.com/appletea233/LLaVA-ST .

We introduce TimeViper, a hybrid vision-language model designed to tackle challenges of long video understanding. Processing long videos demands both an efficient model architecture and an effective mechanism for handling extended temporal contexts. To this end, TimeViper adopts a hybrid Mamba-Transformer backbone that combines the efficiency of state-space models with the expressivity of attention mechanisms. Through this hybrid design, we reveal the vision-to-text information aggregation phenomenon, where information progressively flows from vision tokens to text tokens across increasing LLM depth, resulting in severe vision token redundancy. Motivated by this observation, we propose TransV, a token information transfer module that transfers and compresses vision tokens into instruction tokens while maintaining multimodal understanding capabilities. This design enables TimeViper to process hour-long videos exceeding 10,000 frames. Extensive experiments across multiple benchmarks demonstrate that TimeViper competes with state-of-the-art models while extending frame numbers. We further analyze attention behaviors of both Mamba and Transformer layers, offering new insights into hybrid model interpretability. This work represents an initial step towards developing, interpreting, and compressing hybrid Mamba-Transformer architectures.

Grounding language queries in videos aims at identifying the time interval (or moment) semantically relevant to a language query. The solution to this challenging task demands understanding videos' and queries' semantic content and the fine-grained reasoning about their multi-modal interactions. Our key idea is to recast this challenge into an algorithmic graph matching problem. Fueled by recent advances in Graph Neural Networks, we propose to leverage Graph Convolutional Networks to model video and textual information as well as their semantic alignment. To enable the mutual exchange of information across the modalities, we design a novel Video-Language Graph Matching Network (VLG-Net) to match video and query graphs. Core ingredients include representation graphs built atop video snippets and query tokens separately and used to model intra-modality relationships. A Graph Matching layer is adopted for cross-modal context modeling and multi-modal fusion. Finally, moment candidates are created using masked moment attention pooling by fusing the moment's enriched snippet features. We demonstrate superior performance over state-of-the-art grounding methods on three widely used datasets for temporal localization of moments in videos with language queries: ActivityNet-Captions, TACoS, and DiDeMo.

Understanding multimodal signals in egocentric vision, such as RGB video, depth, camera poses, and gaze, is essential for applications in augmented reality, robotics, and human-computer interaction, enabling systems to better interpret the camera wearer's actions, intentions, and surrounding environment. However, building large-scale egocentric multimodal and multitask models presents unique challenges. Egocentric data are inherently heterogeneous, with large variations in modality coverage across devices and settings. Generating pseudo-labels for missing modalities, such as gaze or head-mounted camera trajectories, is often infeasible, making standard supervised learning approaches difficult to scale. Furthermore, dynamic camera motion and the complex temporal and spatial structure of first-person video pose additional challenges for the direct application of existing multimodal foundation models. To address these challenges, we introduce a set of efficient temporal tokenizers and propose EgoM2P, a masked modeling framework that learns from temporally-aware multimodal tokens to train a large, general-purpose model for egocentric 4D understanding. This unified design supports multitasking across diverse egocentric perception and synthesis tasks, including gaze prediction, egocentric camera tracking, and monocular depth estimation from egocentric video, and also serves as a generative model for conditional egocentric video synthesis. Across these tasks, EgoM2P matches or outperforms specialist models while being an order of magnitude faster. We will fully open-source EgoM2P to support the community and advance egocentric vision research. Project page: https://egom2p.github.io/.

The ability to integrate context, including perceptual and temporal cues, plays a pivotal role in grounding the meaning of a linguistic utterance. In order to measure to what extent current vision-and-language models master this ability, we devise a new multimodal challenge, Image Retrieval from Contextual Descriptions (ImageCoDe). In particular, models are tasked with retrieving the correct image from a set of 10 minimally contrastive candidates based on a contextual description. As such, each description contains only the details that help distinguish between images. Because of this, descriptions tend to be complex in terms of syntax and discourse and require drawing pragmatic inferences. Images are sourced from both static pictures and video frames. We benchmark several state-of-the-art models, including both cross-encoders such as ViLBERT and bi-encoders such as CLIP, on ImageCoDe. Our results reveal that these models dramatically lag behind human performance: the best variant achieves an accuracy of 20.9 on video frames and 59.4 on static pictures, compared with 90.8 in humans. Furthermore, we experiment with new model variants that are better equipped to incorporate visual and temporal context into their representations, which achieve modest gains. Our hope is that ImageCoDE will foster progress in grounded language understanding by encouraging models to focus on fine-grained visual differences.

Recently, large-scale pre-trained language-image models like CLIP have shown extraordinary capabilities for understanding spatial contents, but naively transferring such models to video recognition still suffers from unsatisfactory temporal modeling capabilities. Existing methods insert tunable structures into or in parallel with the pre-trained model, which either requires back-propagation through the whole pre-trained model and is thus resource-demanding, or is limited by the temporal reasoning capability of the pre-trained structure. In this work, we present DiST, which disentangles the learning of spatial and temporal aspects of videos. Specifically, DiST uses a dual-encoder structure, where a pre-trained foundation model acts as the spatial encoder, and a lightweight network is introduced as the temporal encoder. An integration branch is inserted between the encoders to fuse spatio-temporal information. The disentangled spatial and temporal learning in DiST is highly efficient because it avoids the back-propagation of massive pre-trained parameters. Meanwhile, we empirically show that disentangled learning with an extra network for integration benefits both spatial and temporal understanding. Extensive experiments on five benchmarks show that DiST delivers better performance than existing state-of-the-art methods by convincing gaps. When pre-training on the large-scale Kinetics-710, we achieve 89.7% on Kinetics-400 with a frozen ViT-L model, which verifies the scalability of DiST. Codes and models can be found in https://github.com/alibaba-mmai-research/DiST.

We simplify space binding by focusing on two core components, a single encoder per modality and high-quality data; enabling training state-of-the-art models on a single GPU in a few hours as opposed to multiple days. We present EBind, an Easy, data-centric, and parameter-efficient method to Bind the embedding spaces of multiple contrastive models. We demonstrate that a simple 1.8B-parameter image-text-video-audio-3D model can outperform models 4 to 17x the size. The key to achieving this is a carefully curated dataset of three complementary data sources: i) 6.7M fully-automated multimodal quintuples sourced via SOTA retrieval models, ii) 1M diverse, semi-automated triples annotated by humans as negative, partial, or positive matches, and iii) 3.4M pre-existing captioned data items. We use 13 different evaluations to demonstrate the value of each data source. Due to limitations with existing benchmarks, we further introduce the first high-quality, consensus-annotated zero-shot classification benchmark between audio and PCs. In contrast to related work, we will open-source our code, model weights, and datasets.

Vision and language understanding has emerged as a subject undergoing intense study in Artificial Intelligence. Among many tasks in this line of research, visual question answering (VQA) has been one of the most successful ones, where the goal is to learn a model that understands visual content at region-level details and finds their associations with pairs of questions and answers in the natural language form. Despite the rapid progress in the past few years, most existing work in VQA have focused primarily on images. In this paper, we focus on extending VQA to the video domain and contribute to the literature in three important ways. First, we propose three new tasks designed specifically for video VQA, which require spatio-temporal reasoning from videos to answer questions correctly. Next, we introduce a new large-scale dataset for video VQA named TGIF-QA that extends existing VQA work with our new tasks. Finally, we propose a dual-LSTM based approach with both spatial and temporal attention, and show its effectiveness over conventional VQA techniques through empirical evaluations.

Despite advances in general video understanding, Video Large Language Models (Video-LLMs) face challenges in precise temporal localization due to discrete time representations and limited temporally aware datasets. Existing methods for temporal expression either conflate time with text-based numerical values, add a series of dedicated temporal tokens, or regress time using specialized temporal grounding heads. To address these issues, we introduce DisTime, a lightweight framework designed to enhance temporal comprehension in Video-LLMs. DisTime employs a learnable token to create a continuous temporal embedding space and incorporates a Distribution-based Time Decoder that generates temporal probability distributions, effectively mitigating boundary ambiguities and maintaining temporal continuity. Additionally, the Distribution-based Time Encoder re-encodes timestamps to provide time markers for Video-LLMs. To overcome temporal granularity limitations in existing datasets, we propose an automated annotation paradigm that combines the captioning capabilities of Video-LLMs with the localization expertise of dedicated temporal models. This leads to the creation of InternVid-TG, a substantial dataset with 1.25M temporally grounded events across 179k videos, surpassing ActivityNet-Caption by 55 times. Extensive experiments demonstrate that DisTime achieves state-of-the-art performance across benchmarks in three time-sensitive tasks while maintaining competitive performance in Video QA tasks. Code and data are released at https://github.com/josephzpng/DisTime.

Large vision-language models (LVLMs) achieve strong performance on multimodal tasks, yet they often default to their language prior (LP) -- memorized textual patterns from pre-training while under-utilizing visual evidence. Prior analyses of LP mostly rely on input-output probing, which fails to reveal the internal mechanisms governing when and how vision influences model behavior. To address this gap, we present the first systematic analysis of language prior through the lens of chain-of-embedding, which examines the layer-wise representation dynamics within LVLMs. Our analysis reveals a universal phenomenon: each model exhibits a Visual Integration Point (VIP), a critical layer at which visual information begins to meaningfully reshape hidden representations and influence decoding. Building on this observation, we introduce the Total Visual Integration (TVI) estimator, which aggregates representation distance beyond the VIP to quantify how strongly visual query influences response generation. Across 54 model-dataset combinations spanning 9 contemporary LVLMs and 6 benchmarks, we demonstrate that VIP consistently emerges, and that TVI reliably predicts the strength of language prior. This offers a principled toolkit for diagnosing and understanding language prior in LVLMs.

Understanding fine-grained temporal dynamics is crucial for multimodal video comprehension and generation. Due to the lack of fine-grained temporal annotations, existing video benchmarks mostly resemble static image benchmarks and are incompetent at evaluating models for temporal understanding. In this paper, we introduce TemporalBench, a new benchmark dedicated to evaluating fine-grained temporal understanding in videos. TemporalBench consists of ~10K video question-answer pairs, derived from ~2K high-quality human annotations detailing the temporal dynamics in video clips. As a result, our benchmark provides a unique testbed for evaluating various temporal understanding and reasoning abilities such as action frequency, motion magnitude, event order, etc. Moreover, it enables evaluations on various tasks like both video question answering and captioning, both short and long video understanding, as well as different models such as multimodal video embedding models and text generation models. Results show that state-of-the-art models like GPT-4o achieve only 38.5% question answering accuracy on TemporalBench, demonstrating a significant gap (~30%) between humans and AI in temporal understanding. Furthermore, we notice a critical pitfall for multi-choice QA where LLMs can detect the subtle changes in negative captions and find a centralized description as a cue for its prediction, where we propose Multiple Binary Accuracy (MBA) to correct such bias. We hope that TemporalBench can foster research on improving models' temporal reasoning capabilities. Both dataset and evaluation code will be made available.

Scenes are complex, yet structured collections of parts, including objects and surfaces, that exhibit spatial and semantic relations to one another. An effective visual system therefore needs unified scene representations that relate scene parts to their location and their co-occurrence. We hypothesize that this structure can be learned self-supervised from natural experience by exploiting the temporal regularities of active vision: each fixation reveals a locally-detailed glimpse that is statistically related to the previous one via co-occurrence and saccade-conditioned spatial regularities. We instantiate this idea with Glimpse Prediction Networks (GPNs) -- recurrent models trained to predict the feature embedding of the next glimpse along human-like scanpaths over natural scenes. GPNs successfully learn co-occurrence structure and, when given relative saccade location vectors, show sensitivity to spatial arrangement. Furthermore, recurrent variants of GPNs were able to integrate information across glimpses into a unified scene representation. Notably, these scene representations align strongly with human fMRI responses during natural-scene viewing across mid/high-level visual cortex. Critically, GPNs outperform architecture- and dataset-matched controls trained with explicit semantic objectives, and match or exceed strong modern vision baselines, leaving little unique variance for those alternatives. These results establish next-glimpse prediction during active vision as a biologically plausible, self-supervised route to brain-aligned scene representations learned from natural visual experience.

Contrastive language-image pretraining has shown great success in learning visual-textual joint representation from web-scale data, demonstrating remarkable "zero-shot" generalization ability for various image tasks. However, how to effectively expand such new language-image pretraining methods to video domains is still an open problem. In this work, we present a simple yet effective approach that adapts the pretrained language-image models to video recognition directly, instead of pretraining a new model from scratch. More concretely, to capture the long-range dependencies of frames along the temporal dimension, we propose a cross-frame attention mechanism that explicitly exchanges information across frames. Such module is lightweight and can be plugged into pretrained language-image models seamlessly. Moreover, we propose a video-specific prompting scheme, which leverages video content information for generating discriminative textual prompts. Extensive experiments demonstrate that our approach is effective and can be generalized to different video recognition scenarios. In particular, under fully-supervised settings, our approach achieves a top-1 accuracy of 87.1% on Kinectics-400, while using 12 times fewer FLOPs compared with Swin-L and ViViT-H. In zero-shot experiments, our approach surpasses the current state-of-the-art methods by +7.6% and +14.9% in terms of top-1 accuracy under two popular protocols. In few-shot scenarios, our approach outperforms previous best methods by +32.1% and +23.1% when the labeled data is extremely limited. Code and models are available at https://aka.ms/X-CLIP

Spatio-temporal grounding describes the task of localizing events in space and time, e.g., in video data, based on verbal descriptions only. Models for this task are usually trained with human-annotated sentences and bounding box supervision. This work addresses this task from a multimodal supervision perspective, proposing a framework for spatio-temporal action grounding trained on loose video and subtitle supervision only, without human annotation. To this end, we combine local representation learning, which focuses on leveraging fine-grained spatial information, with a global representation encoding that captures higher-level representations and incorporates both in a joint approach. To evaluate this challenging task in a real-life setting, a new benchmark dataset is proposed providing dense spatio-temporal grounding annotations in long, untrimmed, multi-action instructional videos for over 5K events. We evaluate the proposed approach and other methods on the proposed and standard downstream tasks showing that our method improves over current baselines in various settings, including spatial, temporal, and untrimmed multi-action spatio-temporal grounding.

Visual tempo, which describes how fast an action goes, has shown its potential in supervised action recognition. In this work, we demonstrate that visual tempo can also serve as a self-supervision signal for video representation learning. We propose to maximize the mutual information between representations of slow and fast videos via hierarchical contrastive learning (VTHCL). Specifically, by sampling the same instance at slow and fast frame rates respectively, we can obtain slow and fast video frames which share the same semantics but contain different visual tempos. Video representations learned from VTHCL achieve the competitive performances under the self-supervision evaluation protocol for action recognition on UCF-101 (82.1\%) and HMDB-51 (49.2\%). Moreover, comprehensive experiments suggest that the learned representations are generalized well to other downstream tasks including action detection on AVA and action anticipation on Epic-Kitchen. Finally, we propose Instance Correspondence Map (ICM) to visualize the shared semantics captured by contrastive learning.

The integration of Large Language Models (LLMs) with visual encoders has recently shown promising performance in visual understanding tasks, leveraging their inherent capability to comprehend and generate human-like text for visual reasoning. Given the diverse nature of visual data, MultiModal Large Language Models (MM-LLMs) exhibit variations in model designing and training for understanding images, short videos, and long videos. Our paper focuses on the substantial differences and unique challenges posed by long video understanding compared to static image and short video understanding. Unlike static images, short videos encompass sequential frames with both spatial and within-event temporal information, while long videos consist of multiple events with between-event and long-term temporal information. In this survey, we aim to trace and summarize the advancements of MM-LLMs from image understanding to long video understanding. We review the differences among various visual understanding tasks and highlight the challenges in long video understanding, including more fine-grained spatiotemporal details, dynamic events, and long-term dependencies. We then provide a detailed summary of the advancements in MM-LLMs in terms of model design and training methodologies for understanding long videos. Finally, we compare the performance of existing MM-LLMs on video understanding benchmarks of various lengths and discuss potential future directions for MM-LLMs in long video understanding.

In this technical report, we introduce our solution to human-centric spatio-temporal video grounding task. We propose a concise and effective framework named STVGFormer, which models spatiotemporal visual-linguistic dependencies with a static branch and a dynamic branch. The static branch performs cross-modal understanding in a single frame and learns to localize the target object spatially according to intra-frame visual cues like object appearances. The dynamic branch performs cross-modal understanding across multiple frames. It learns to predict the starting and ending time of the target moment according to dynamic visual cues like motions. Both the static and dynamic branches are designed as cross-modal transformers. We further design a novel static-dynamic interaction block to enable the static and dynamic branches to transfer useful and complementary information from each other, which is shown to be effective to improve the prediction on hard cases. Our proposed method achieved 39.6% vIoU and won the first place in the HC-STVG track of the 4th Person in Context Challenge.

In this paper, we study the problem of temporal video grounding (TVG), which aims to predict the starting/ending time points of moments described by a text sentence within a long untrimmed video. Benefiting from fine-grained 3D visual features, the TVG techniques have achieved remarkable progress in recent years. However, the high complexity of 3D convolutional neural networks (CNNs) makes extracting dense 3D visual features time-consuming, which calls for intensive memory and computing resources. Towards efficient TVG, we propose a novel text-visual prompting (TVP) framework, which incorporates optimized perturbation patterns (that we call 'prompts') into both visual inputs and textual features of a TVG model. In sharp contrast to 3D CNNs, we show that TVP allows us to effectively co-train vision encoder and language encoder in a 2D TVG model and improves the performance of crossmodal feature fusion using only low-complexity sparse 2D visual features. Further, we propose a Temporal-Distance IoU (TDIoU) loss for efficient learning of TVG. Experiments on two benchmark datasets, Charades-STA and ActivityNet Captions datasets, empirically show that the proposed TVP significantly boosts the performance of 2D TVG (e.g., 9.79% improvement on Charades-STA and 30.77% improvement on ActivityNet Captions) and achieves 5x inference acceleration over TVG using 3D visual features. Codes are available at Open.Intel.

We use multilayer Long Short Term Memory (LSTM) networks to learn representations of video sequences. Our model uses an encoder LSTM to map an input sequence into a fixed length representation. This representation is decoded using single or multiple decoder LSTMs to perform different tasks, such as reconstructing the input sequence, or predicting the future sequence. We experiment with two kinds of input sequences - patches of image pixels and high-level representations ("percepts") of video frames extracted using a pretrained convolutional net. We explore different design choices such as whether the decoder LSTMs should condition on the generated output. We analyze the outputs of the model qualitatively to see how well the model can extrapolate the learned video representation into the future and into the past. We try to visualize and interpret the learned features. We stress test the model by running it on longer time scales and on out-of-domain data. We further evaluate the representations by finetuning them for a supervised learning problem - human action recognition on the UCF-101 and HMDB-51 datasets. We show that the representations help improve classification accuracy, especially when there are only a few training examples. Even models pretrained on unrelated datasets (300 hours of YouTube videos) can help action recognition performance.

Zero-shot learning (ZSL) models rely on learning a joint embedding space where both textual/semantic description of object classes and visual representation of object images can be projected to for nearest neighbour search. Despite the success of deep neural networks that learn an end-to-end model between text and images in other vision problems such as image captioning, very few deep ZSL model exists and they show little advantage over ZSL models that utilise deep feature representations but do not learn an end-to-end embedding. In this paper we argue that the key to make deep ZSL models succeed is to choose the right embedding space. Instead of embedding into a semantic space or an intermediate space, we propose to use the visual space as the embedding space. This is because that in this space, the subsequent nearest neighbour search would suffer much less from the hubness problem and thus become more effective. This model design also provides a natural mechanism for multiple semantic modalities (e.g., attributes and sentence descriptions) to be fused and optimised jointly in an end-to-end manner. Extensive experiments on four benchmarks show that our model significantly outperforms the existing models. Code is available at https://github.com/lzrobots/DeepEmbeddingModel_ZSL

We study the intriguing connection between visual data, deep networks, and the brain. Our method creates a universal channel alignment by using brain voxel fMRI response prediction as the training objective. We discover that deep networks, trained with different objectives, share common feature channels across various models. These channels can be clustered into recurring sets, corresponding to distinct brain regions, indicating the formation of visual concepts. Tracing the clusters of channel responses onto the images, we see semantically meaningful object segments emerge, even without any supervised decoder. Furthermore, the universal feature alignment and the clustering of channels produce a picture and quantification of how visual information is processed through the different network layers, which produces precise comparisons between the networks.

We present ImageBind, an approach to learn a joint embedding across six different modalities - images, text, audio, depth, thermal, and IMU data. We show that all combinations of paired data are not necessary to train such a joint embedding, and only image-paired data is sufficient to bind the modalities together. ImageBind can leverage recent large scale vision-language models, and extends their zero-shot capabilities to new modalities just by using their natural pairing with images. It enables novel emergent applications 'out-of-the-box' including cross-modal retrieval, composing modalities with arithmetic, cross-modal detection and generation. The emergent capabilities improve with the strength of the image encoder and we set a new state-of-the-art on emergent zero-shot recognition tasks across modalities, outperforming specialist supervised models. Finally, we show strong few-shot recognition results outperforming prior work, and that ImageBind serves as a new way to evaluate vision models for visual and non-visual tasks.

Leveraging temporal information has been regarded as essential for developing video understanding models. However, how to properly incorporate temporal information into the recent successful instance discrimination based contrastive self-supervised learning (CSL) framework remains unclear. As an intuitive solution, we find that directly applying temporal augmentations does not help, or even impair video CSL in general. This counter-intuitive observation motivates us to re-design existing video CSL frameworks, for better integration of temporal knowledge. To this end, we present Temporal-aware Contrastive self-supervised learningTaCo, as a general paradigm to enhance video CSL. Specifically, TaCo selects a set of temporal transformations not only as strong data augmentation but also to constitute extra self-supervision for video understanding. By jointly contrasting instances with enriched temporal transformations and learning these transformations as self-supervised signals, TaCo can significantly enhance unsupervised video representation learning. For instance, TaCo demonstrates consistent improvement in downstream classification tasks over a list of backbones and CSL approaches. Our best model achieves 85.1% (UCF-101) and 51.6% (HMDB-51) top-1 accuracy, which is a 3% and 2.4% relative improvement over the previous state-of-the-art.

Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) stand as the two most popular foundation models for visual representation learning. While CNNs exhibit remarkable scalability with linear complexity w.r.t. image resolution, ViTs surpass them in fitting capabilities despite contending with quadratic complexity. A closer inspection reveals that ViTs achieve superior visual modeling performance through the incorporation of global receptive fields and dynamic weights. This observation motivates us to propose a novel architecture that inherits these components while enhancing computational efficiency. To this end, we draw inspiration from the recently introduced state space model and propose the Visual State Space Model (VMamba), which achieves linear complexity without sacrificing global receptive fields. To address the encountered direction-sensitive issue, we introduce the Cross-Scan Module (CSM) to traverse the spatial domain and convert any non-causal visual image into order patch sequences. Extensive experimental results substantiate that VMamba not only demonstrates promising capabilities across various visual perception tasks, but also exhibits more pronounced advantages over established benchmarks as the image resolution increases. Source code has been available at https://github.com/MzeroMiko/VMamba.

Temporal action detection (TAD) involves the localization and classification of action instances within untrimmed videos. While standard TAD follows fully supervised learning with closed-set setting on large training data, recent zero-shot TAD methods showcase the promising open-set setting by leveraging large-scale contrastive visual-language (ViL) pretrained models. However, existing zero-shot TAD methods have limitations on how to properly construct the strong relationship between two interdependent tasks of localization and classification and adapt ViL model to video understanding. In this work, we present ZEETAD, featuring two modules: dual-localization and zero-shot proposal classification. The former is a Transformer-based module that detects action events while selectively collecting crucial semantic embeddings for later recognition. The latter one, CLIP-based module, generates semantic embeddings from text and frame inputs for each temporal unit. Additionally, we enhance discriminative capability on unseen classes by minimally updating the frozen CLIP encoder with lightweight adapters. Extensive experiments on THUMOS14 and ActivityNet-1.3 datasets demonstrate our approach's superior performance in zero-shot TAD and effective knowledge transfer from ViL models to unseen action categories.

This paper presents a simple yet effective approach to modeling space-time correspondences in the context of video object segmentation. Unlike most existing approaches, we establish correspondences directly between frames without re-encoding the mask features for every object, leading to a highly efficient and robust framework. With the correspondences, every node in the current query frame is inferred by aggregating features from the past in an associative fashion. We cast the aggregation process as a voting problem and find that the existing inner-product affinity leads to poor use of memory with a small (fixed) subset of memory nodes dominating the votes, regardless of the query. In light of this phenomenon, we propose using the negative squared Euclidean distance instead to compute the affinities. We validated that every memory node now has a chance to contribute, and experimentally showed that such diversified voting is beneficial to both memory efficiency and inference accuracy. The synergy of correspondence networks and diversified voting works exceedingly well, achieves new state-of-the-art results on both DAVIS and YouTubeVOS datasets while running significantly faster at 20+ FPS for multiple objects without bells and whistles.

We consider the task of temporal human action localization in lifestyle vlogs. We introduce a novel dataset consisting of manual annotations of temporal localization for 13,000 narrated actions in 1,200 video clips. We present an extensive analysis of this data, which allows us to better understand how the language and visual modalities interact throughout the videos. We propose a simple yet effective method to localize the narrated actions based on their expected duration. Through several experiments and analyses, we show that our method brings complementary information with respect to previous methods, and leads to improvements over previous work for the task of temporal action localization.

Human action understanding is crucial for the advancement of multimodal systems. While recent developments, driven by powerful large language models (LLMs), aim to be general enough to cover a wide range of categories, they often overlook the need for more specific capabilities. In this work, we address the more challenging task of Fine-grained Action Recognition (FAR), which focuses on detailed semantic labels within shorter temporal duration (e.g., "salto backward tucked with 1 turn"). Given the high costs of annotating fine-grained labels and the substantial data needed for fine-tuning LLMs, we propose to adopt semi-supervised learning (SSL). Our framework, SeFAR, incorporates several innovative designs to tackle these challenges. Specifically, to capture sufficient visual details, we construct Dual-level temporal elements as more effective representations, based on which we design a new strong augmentation strategy for the Teacher-Student learning paradigm through involving moderate temporal perturbation. Furthermore, to handle the high uncertainty within the teacher model's predictions for FAR, we propose the Adaptive Regulation to stabilize the learning process. Experiments show that SeFAR achieves state-of-the-art performance on two FAR datasets, FineGym and FineDiving, across various data scopes. It also outperforms other semi-supervised methods on two classical coarse-grained datasets, UCF101 and HMDB51. Further analysis and ablation studies validate the effectiveness of our designs. Additionally, we show that the features extracted by our SeFAR could largely promote the ability of multimodal foundation models to understand fine-grained and domain-specific semantics.

Despite significant advances in Multimodal Large Language Models (MLLMs), understanding complex temporal dynamics in videos remains a major challenge. Our experiments show that current Video Large Language Model (Video-LLM) architectures have critical limitations in temporal understanding, struggling with tasks that require detailed comprehension of action sequences and temporal progression. In this work, we propose a Video-LLM architecture that introduces stacked temporal attention modules directly within the vision encoder. This design incorporates a temporal attention in vision encoder, enabling the model to better capture the progression of actions and the relationships between frames before passing visual tokens to the LLM. Our results show that this approach significantly improves temporal reasoning and outperforms existing models in video question answering tasks, specifically in action recognition. We improve on benchmarks including VITATECS, MVBench, and Video-MME by up to +5.5%. By enhancing the vision encoder with temporal structure, we address a critical gap in video understanding for Video-LLMs. Project page and code are available at: https://alirasekh.github.io/STAVEQ2/.

Learning a good representation for space-time correspondence is the key for various computer vision tasks, including tracking object bounding boxes and performing video object pixel segmentation. To learn generalizable representation for correspondence in large-scale, a variety of self-supervised pretext tasks are proposed to explicitly perform object-level or patch-level similarity learning. Instead of following the previous literature, we propose to learn correspondence using Video Frame-level Similarity (VFS) learning, i.e, simply learning from comparing video frames. Our work is inspired by the recent success in image-level contrastive learning and similarity learning for visual recognition. Our hypothesis is that if the representation is good for recognition, it requires the convolutional features to find correspondence between similar objects or parts. Our experiments show surprising results that VFS surpasses state-of-the-art self-supervised approaches for both OTB visual object tracking and DAVIS video object segmentation. We perform detailed analysis on what matters in VFS and reveals new properties on image and frame level similarity learning. Project page with code is available at https://jerryxu.net/VFS

Existing multi-person video pose estimation methods typically adopt a two-stage pipeline: detecting individuals in each frame, followed by temporal modeling for single-person pose estimation. This design relies on heuristic operations such as detection, RoI cropping, and non-maximum suppression (NMS), limiting both accuracy and efficiency. In this paper, we present a fully end-to-end framework for multi-person 2D pose estimation in videos, effectively eliminating heuristic operations. A key challenge is to associate individuals across frames under complex and overlapping temporal trajectories. To address this, we introduce a novel Pose-Aware Video transformEr Network (PAVE-Net), which features a spatial encoder to model intra-frame relations and a spatiotemporal pose decoder to capture global dependencies across frames. To achieve accurate temporal association, we propose a pose-aware attention mechanism that enables each pose query to selectively aggregate features corresponding to the same individual across consecutive frames.Additionally, we explicitly model spatiotemporal dependencies among pose keypoints to improve accuracy. Notably, our approach is the first end-to-end method for multi-frame 2D human pose estimation.Extensive experiments show that PAVE-Net substantially outperforms prior image-based end-to-end methods, achieving a 6.0 mAP improvement on PoseTrack2017, and delivers accuracy competitive with state-of-the-art two-stage video-based approaches, while offering significant gains in efficiency.Project page: https://github.com/zgspose/PAVENet

We present CAT-V (Caption AnyThing in Video), a training-free framework for fine-grained object-centric video captioning that enables detailed descriptions of user-selected objects through time. CAT-V integrates three key components: a Segmenter based on SAMURAI for precise object segmentation across frames, a Temporal Analyzer powered by TRACE-Uni for accurate event boundary detection and temporal analysis, and a Captioner using InternVL-2.5 for generating detailed object-centric descriptions. Through spatiotemporal visual prompts and chain-of-thought reasoning, our framework generates detailed, temporally-aware descriptions of objects' attributes, actions, statuses, interactions, and environmental contexts without requiring additional training data. CAT-V supports flexible user interactions through various visual prompts (points, bounding boxes, and irregular regions) and maintains temporal sensitivity by tracking object states and interactions across different time segments. Our approach addresses limitations of existing video captioning methods, which either produce overly abstract descriptions or lack object-level precision, enabling fine-grained, object-specific descriptions while maintaining temporal coherence and spatial accuracy. The GitHub repository for this project is available at https://github.com/yunlong10/CAT-V

This paper studies a conceptually simple extension of Masked Autoencoders (MAE) to spatiotemporal representation learning from videos. We randomly mask out spacetime patches in videos and learn an autoencoder to reconstruct them in pixels. Interestingly, we show that our MAE method can learn strong representations with almost no inductive bias on spacetime (only except for patch and positional embeddings), and spacetime-agnostic random masking performs the best. We observe that the optimal masking ratio is as high as 90% (vs. 75% on images), supporting the hypothesis that this ratio is related to information redundancy of the data. A high masking ratio leads to a large speedup, e.g., > 4x in wall-clock time or even more. We report competitive results on several challenging video datasets using vanilla Vision Transformers. We observe that MAE can outperform supervised pre-training by large margins. We further report encouraging results of training on real-world, uncurated Instagram data. Our study suggests that the general framework of masked autoencoding (BERT, MAE, etc.) can be a unified methodology for representation learning with minimal domain knowledge.

Existing video-language pre-training methods primarily focus on instance-level alignment between video clips and captions via global contrastive learning but neglect rich fine-grained local information in both videos and text, which is of importance to downstream tasks requiring temporal localization and semantic reasoning. A powerful model is expected to be capable of capturing region-object correspondences and recognizing scene changes in a video clip, reflecting spatial and temporal granularity, respectively. To strengthen model's understanding into such fine-grained details, we propose a simple yet effective video-language modeling framework, S-ViLM, by exploiting the intrinsic structures of these two modalities. It includes two novel designs, inter-clip spatial grounding and intra-clip temporal grouping, to promote learning region-object alignment and temporal-aware features, simultaneously. Comprehensive evaluations demonstrate that S-ViLM performs favorably against existing approaches in learning more expressive representations. Specifically, S-ViLM surpasses the state-of-the-art methods substantially on four representative downstream tasks, covering text-video retrieval, video question answering, video action recognition, and temporal action localization.

Current video-based Masked Autoencoders (MAEs) primarily focus on learning effective spatiotemporal representations from a visual perspective, which may lead the model to prioritize general spatial-temporal patterns but often overlook nuanced semantic attributes like specific interactions or sequences that define actions - such as action-specific features that align more closely with human cognition for space-time correspondence. This can limit the model's ability to capture the essence of certain actions that are contextually rich and continuous. Humans are capable of mapping visual concepts, object view invariance, and semantic attributes available in static instances to comprehend natural dynamic scenes or videos. Existing MAEs for videos and static images rely on separate datasets for videos and images, which may lack the rich semantic attributes necessary for fully understanding the learned concepts, especially when compared to using video and corresponding sampled frame images together. To this end, we propose CrossVideoMAE an end-to-end self-supervised cross-modal contrastive learning MAE that effectively learns both video-level and frame-level rich spatiotemporal representations and semantic attributes. Our method integrates mutual spatiotemporal information from videos with spatial information from sampled frames within a feature-invariant space, while encouraging invariance to augmentations within the video domain. This objective is achieved through jointly embedding features of visible tokens and combining feature correspondence within and across modalities, which is critical for acquiring rich, label-free guiding signals from both video and frame image modalities in a self-supervised manner. Extensive experiments demonstrate that our approach surpasses previous state-of-the-art methods and ablation studies validate the effectiveness of our approach.

The joint-embedding predictive architecture (JEPA) recently has shown impressive results in extracting visual representations from unlabeled imagery under a masking strategy. However, we reveal its disadvantages, notably its insufficient understanding of local semantics. This deficiency originates from masked modeling in the embedding space, resulting in a reduction of discriminative power and can even lead to the neglect of critical local semantics. To bridge this gap, we introduce DMT-JEPA, a novel masked modeling objective rooted in JEPA, specifically designed to generate discriminative latent targets from neighboring information. Our key idea is simple: we consider a set of semantically similar neighboring patches as a target of a masked patch. To be specific, the proposed DMT-JEPA (a) computes feature similarities between each masked patch and its corresponding neighboring patches to select patches having semantically meaningful relations, and (b) employs lightweight cross-attention heads to aggregate features of neighboring patches as the masked targets. Consequently, DMT-JEPA demonstrates strong discriminative power, offering benefits across a diverse spectrum of downstream tasks. Through extensive experiments, we demonstrate our effectiveness across various visual benchmarks, including ImageNet-1K image classification, ADE20K semantic segmentation, and COCO object detection tasks. Code is available at: https://github.com/DMTJEPA/DMTJEPA.

We have witnessed impressive advances in video action understanding. Increased dataset sizes, variability, and computation availability have enabled leaps in performance and task diversification. Current systems can provide coarse- and fine-grained descriptions of video scenes, extract segments corresponding to queries, synthesize unobserved parts of videos, and predict context. This survey comprehensively reviews advances in uni- and multi-modal action understanding across a range of tasks. We focus on prevalent challenges, overview widely adopted datasets, and survey seminal works with an emphasis on recent advances. We broadly distinguish between three temporal scopes: (1) recognition tasks of actions observed in full, (2) prediction tasks for ongoing partially observed actions, and (3) forecasting tasks for subsequent unobserved action. This division allows us to identify specific action modeling and video representation challenges. Finally, we outline future directions to address current shortcomings.

Video-language pre-training has advanced the performance of various downstream video-language tasks. However, most previous methods directly inherit or adapt typical image-language pre-training paradigms to video-language pre-training, thus not fully exploiting the unique characteristic of video, i.e., temporal. In this paper, we propose a Hierarchical Temporal-Aware video-language pre-training framework, HiTeA, with two novel pre-training tasks for modeling cross-modal alignment between moments and texts as well as the temporal relations of video-text pairs. Specifically, we propose a cross-modal moment exploration task to explore moments in videos, which results in detailed video moment representation. Besides, the inherent temporal relations are captured by aligning video-text pairs as a whole in different time resolutions with multi-modal temporal relation exploration task. Furthermore, we introduce the shuffling test to evaluate the temporal reliance of datasets and video-language pre-training models. We achieve state-of-the-art results on 15 well-established video-language understanding and generation tasks, especially on temporal-oriented datasets (e.g., SSv2-Template and SSv2-Label) with 8.6% and 11.1% improvement respectively. HiTeA also demonstrates strong generalization ability when directly transferred to downstream tasks in a zero-shot manner. Models and demo will be available on ModelScope.

Many studies in vision tasks have aimed to create effective embedding spaces for single-label object prediction within an image. However, in reality, most objects possess multiple specific attributes, such as shape, color, and length, with each attribute composed of various classes. To apply models in real-world scenarios, it is essential to be able to distinguish between the granular components of an object. Conventional approaches to embedding multiple specific attributes into a single network often result in entanglement, where fine-grained features of each attribute cannot be identified separately. To address this problem, we propose a Conditional Cross-Attention Network that induces disentangled multi-space embeddings for various specific attributes with only a single backbone. Firstly, we employ a cross-attention mechanism to fuse and switch the information of conditions (specific attributes), and we demonstrate its effectiveness through a diverse visualization example. Secondly, we leverage the vision transformer for the first time to a fine-grained image retrieval task and present a simple yet effective framework compared to existing methods. Unlike previous studies where performance varied depending on the benchmark dataset, our proposed method achieved consistent state-of-the-art performance on the FashionAI, DARN, DeepFashion, and Zappos50K benchmark datasets.

Temporal sentence grounding in videos (TSGV), \aka natural language video localization (NLVL) or video moment retrieval (VMR), aims to retrieve a temporal moment that semantically corresponds to a language query from an untrimmed video. Connecting computer vision and natural language, TSGV has drawn significant attention from researchers in both communities. This survey attempts to provide a summary of fundamental concepts in TSGV and current research status, as well as future research directions. As the background, we present a common structure of functional components in TSGV, in a tutorial style: from feature extraction from raw video and language query, to answer prediction of the target moment. Then we review the techniques for multimodal understanding and interaction, which is the key focus of TSGV for effective alignment between the two modalities. We construct a taxonomy of TSGV techniques and elaborate the methods in different categories with their strengths and weaknesses. Lastly, we discuss issues with the current TSGV research and share our insights about promising research directions.

Recently, Vision Large Language Models (VLLMs) integrated with vision encoders have shown promising performance in vision understanding. The key of VLLMs is to encode visual content into sequences of visual tokens, enabling VLLMs to simultaneously process both visual and textual content. However, understanding videos, especially long videos, remain a challenge to VLLMs as the number of visual tokens grows rapidly when encoding videos, resulting in the risk of exceeding the context window of VLLMs and introducing heavy computation burden. To restrict the number of visual tokens, existing VLLMs either: (1) uniformly downsample videos into a fixed number of frames or (2) reducing the number of visual tokens encoded from each frame. We argue the former solution neglects the rich temporal cue in videos and the later overlooks the spatial details in each frame. In this work, we present Balanced-VLLM (B-VLLM): a novel VLLM framework that aims to effectively leverage task relevant spatio-temporal cues while restricting the number of visual tokens under the VLLM context window length. At the core of our method, we devise a text-conditioned adaptive frame selection module to identify frames relevant to the visual understanding task. The selected frames are then de-duplicated using a temporal frame token merging technique. The visual tokens of the selected frames are processed through a spatial token sampling module and an optional spatial token merging strategy to achieve precise control over the token count. Experimental results show that B-VLLM is effective in balancing the number of frames and visual tokens in video understanding, yielding superior performance on various video understanding benchmarks. Our code is available at https://github.com/zhuqiangLu/B-VLLM.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However, the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent multimodal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors, we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify ''CLIP-blind pairs'' - images that CLIP perceives as similar despite their clear visual differences. With these pairs, we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems, including GPT-4V, struggle with straightforward questions across nine basic visual patterns, often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues, we propose a Mixture of Features (MoF) approach, demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together, our research suggests visual representation learning remains an open challenge, and accurate visual grounding is crucial for future successful multimodal systems.

Over the past decade, predictive modeling of neural responses in the primate visual system has advanced significantly, largely driven by various DNN approaches. These include models optimized directly for visual recognition, cross-modal alignment through contrastive objectives, neural response prediction from scratch, and large language model embeddings.Likewise, different readout mechanisms, ranging from fully linear to spatial-feature factorized methods have been explored for mapping network activations to neural responses. Despite the diversity of these approaches, it remains unclear which method performs best across different visual regions. In this study, we systematically compare these approaches for modeling the human visual system and investigate alternative strategies to improve response predictions. Our findings reveal that for early to mid-level visual areas, response-optimized models with visual inputs offer superior prediction accuracy, while for higher visual regions, embeddings from LLMs based on detailed contextual descriptions of images and task-optimized models pretrained on large vision datasets provide the best fit. Through comparative analysis of these modeling approaches, we identified three distinct regions in the visual cortex: one sensitive primarily to perceptual features of the input that are not captured by linguistic descriptions, another attuned to fine-grained visual details representing semantic information, and a third responsive to abstract, global meanings aligned with linguistic content. We also highlight the critical role of readout mechanisms, proposing a novel scheme that modulates receptive fields and feature maps based on semantic content, resulting in an accuracy boost of 3-23% over existing SOTAs for all models and brain regions. Together, these findings offer key insights into building more precise models of the visual system.

Current approaches to learning semantic representations of sentences often use prior word-level knowledge. The current study aims to leverage visual information in order to capture sentence level semantics without the need for word embeddings. We use a multimodal sentence encoder trained on a corpus of images with matching text captions to produce visually grounded sentence embeddings. Deep Neural Networks are trained to map the two modalities to a common embedding space such that for an image the corresponding caption can be retrieved and vice versa. We show that our model achieves results comparable to the current state-of-the-art on two popular image-caption retrieval benchmark data sets: MSCOCO and Flickr8k. We evaluate the semantic content of the resulting sentence embeddings using the data from the Semantic Textual Similarity benchmark task and show that the multimodal embeddings correlate well with human semantic similarity judgements. The system achieves state-of-the-art results on several of these benchmarks, which shows that a system trained solely on multimodal data, without assuming any word representations, is able to capture sentence level semantics. Importantly, this result shows that we do not need prior knowledge of lexical level semantics in order to model sentence level semantics. These findings demonstrate the importance of visual information in semantics.

We present a novel technique for self-supervised video representation learning by: (a) decoupling the learning objective into two contrastive subtasks respectively emphasizing spatial and temporal features, and (b) performing it hierarchically to encourage multi-scale understanding. Motivated by their effectiveness in supervised learning, we first introduce spatial-temporal feature learning decoupling and hierarchical learning to the context of unsupervised video learning. We show by experiments that augmentations can be manipulated as regularization to guide the network to learn desired semantics in contrastive learning, and we propose a way for the model to separately capture spatial and temporal features at multiple scales. We also introduce an approach to overcome the problem of divergent levels of instance invariance at different hierarchies by modeling the invariance as loss weights for objective re-weighting. Experiments on downstream action recognition benchmarks on UCF101 and HMDB51 show that our proposed Hierarchically Decoupled Spatial-Temporal Contrast (HDC) makes substantial improvements over directly learning spatial-temporal features as a whole and achieves competitive performance when compared with other state-of-the-art unsupervised methods. Code will be made available.

This work aims to learn a high-quality text-to-video (T2V) generative model by leveraging a pre-trained text-to-image (T2I) model as a basis. It is a highly desirable yet challenging task to simultaneously a) accomplish the synthesis of visually realistic and temporally coherent videos while b) preserving the strong creative generation nature of the pre-trained T2I model. To this end, we propose LaVie, an integrated video generation framework that operates on cascaded video latent diffusion models, comprising a base T2V model, a temporal interpolation model, and a video super-resolution model. Our key insights are two-fold: 1) We reveal that the incorporation of simple temporal self-attentions, coupled with rotary positional encoding, adequately captures the temporal correlations inherent in video data. 2) Additionally, we validate that the process of joint image-video fine-tuning plays a pivotal role in producing high-quality and creative outcomes. To enhance the performance of LaVie, we contribute a comprehensive and diverse video dataset named Vimeo25M, consisting of 25 million text-video pairs that prioritize quality, diversity, and aesthetic appeal. Extensive experiments demonstrate that LaVie achieves state-of-the-art performance both quantitatively and qualitatively. Furthermore, we showcase the versatility of pre-trained LaVie models in various long video generation and personalized video synthesis applications.

Representation learning promises to unlock deep learning for the long tail of vision tasks without expensive labelled datasets. Yet, the absence of a unified evaluation for general visual representations hinders progress. Popular protocols are often too constrained (linear classification), limited in diversity (ImageNet, CIFAR, Pascal-VOC), or only weakly related to representation quality (ELBO, reconstruction error). We present the Visual Task Adaptation Benchmark (VTAB), which defines good representations as those that adapt to diverse, unseen tasks with few examples. With VTAB, we conduct a large-scale study of many popular publicly-available representation learning algorithms. We carefully control confounders such as architecture and tuning budget. We address questions like: How effective are ImageNet representations beyond standard natural datasets? How do representations trained via generative and discriminative models compare? To what extent can self-supervision replace labels? And, how close are we to general visual representations?

Joint embedding spaces have significantly advanced music understanding and generation by linking text and audio through multimodal contrastive learning. However, these approaches face large memory requirement limitations due to relying on large batch sizes to effectively utilize negative samples. Further, multimodal joint embedding spaces suffer from a modality gap wherein embeddings from different modalities lie in different manifolds of the embedding space. To address these challenges, we propose Siamese Language-Audio Pretraining (SLAP), a novel multimodal pretraining framework that allows learning powerful representations without negative samples. SLAP adapts the Bootstrap Your Own Latent (BYOL) paradigm for multimodal audio-text training, promoting scalability in training multimodal embedding spaces. We illustrate the ability of our model to learn meaningful relationships between music and text -- specifically, we show that SLAP outperforms CLAP on tasks such as text-music retrieval and zero-shot classification. We also observe competitive downstream performance on several MIR tasks, including with larger or supervised models (genre and instrument classification, auto-tagging). Additionally, our approach has attractive properties, such as a quantifiably reduced modality gap and improved robustness to batch size variations on retrieval performance. Finally, its novel formulation unlocks large-scale training on a single GPU through gradient accumulation.

We present a meta-learning framework for learning new visual concepts quickly, from just one or a few examples, guided by multiple naturally occurring data streams: simultaneously looking at images, reading sentences that describe the objects in the scene, and interpreting supplemental sentences that relate the novel concept with other concepts. The learned concepts support downstream applications, such as answering questions by reasoning about unseen images. Our model, namely FALCON, represents individual visual concepts, such as colors and shapes, as axis-aligned boxes in a high-dimensional space (the "box embedding space"). Given an input image and its paired sentence, our model first resolves the referential expression in the sentence and associates the novel concept with particular objects in the scene. Next, our model interprets supplemental sentences to relate the novel concept with other known concepts, such as "X has property Y" or "X is a kind of Y". Finally, it infers an optimal box embedding for the novel concept that jointly 1) maximizes the likelihood of the observed instances in the image, and 2) satisfies the relationships between the novel concepts and the known ones. We demonstrate the effectiveness of our model on both synthetic and real-world datasets.

Recently, there is a surge in interest surrounding video large language models (Video LLMs). However, existing benchmarks fail to provide a comprehensive feedback on the temporal perception ability of Video LLMs. On the one hand, most of them are unable to distinguish between different temporal aspects (e.g., speed, direction) and thus cannot reflect the nuanced performance on these specific aspects. On the other hand, they are limited in the diversity of task formats (e.g., only multi-choice QA), which hinders the understanding of how temporal perception performance may vary across different types of tasks. Motivated by these two problems, we propose the TempCompass benchmark, which introduces a diversity of temporal aspects and task formats. To collect high-quality test data, we devise two novel strategies: (1) In video collection, we construct conflicting videos that share the same static content but differ in a specific temporal aspect, which prevents Video LLMs from leveraging single-frame bias or language priors. (2) To collect the task instructions, we propose a paradigm where humans first annotate meta-information for a video and then an LLM generates the instruction. We also design an LLM-based approach to automatically and accurately evaluate the responses from Video LLMs. Based on TempCompass, we comprehensively evaluate 8 state-of-the-art (SOTA) Video LLMs and 3 Image LLMs, and reveal the discerning fact that these models exhibit notably poor temporal perception ability. The data and evaluation code are available at https://github.com/llyx97/TempCompass.

The performance of algorithms for neural architecture search strongly depends on the parametrization of the search space. We use contrastive learning to identify networks across different initializations based on their data Jacobians, and automatically produce the first architecture embeddings independent from the parametrization of the search space. Using our contrastive embeddings, we show that traditional black-box optimization algorithms, without modification, can reach state-of-the-art performance in Neural Architecture Search. As our method provides a unified embedding space, we perform for the first time transfer learning between search spaces. Finally, we show the evolution of embeddings during training, motivating future studies into using embeddings at different training stages to gain a deeper understanding of the networks in a search space.

Unifying diverse image generation tasks within a single framework remains a fundamental challenge in visual generation. While large language models (LLMs) achieve unification through task-agnostic data and generation, existing visual generation models fail to meet these principles. Current approaches either rely on per-task datasets and large-scale training or adapt pre-trained image models with task-specific modifications, limiting their generalizability. In this work, we explore video models as a foundation for unified image generation, leveraging their inherent ability to model temporal correlations. We introduce RealGeneral, a novel framework that reformulates image generation as a conditional frame prediction task, analogous to in-context learning in LLMs. To bridge the gap between video models and condition-image pairs, we propose (1) a Unified Conditional Embedding module for multi-modal alignment and (2) a Unified Stream DiT Block with decoupled adaptive LayerNorm and attention mask to mitigate cross-modal interference. RealGeneral demonstrates effectiveness in multiple important visual generation tasks, e.g., it achieves a 14.5% improvement in subject similarity for customized generation and a 10% enhancement in image quality for canny-to-image task. Project page: https://lyne1.github.io/RealGeneral/

Learning semantically meaningful sentence embeddings is an open problem in natural language processing. In this work, we propose a sentence embedding learning approach that exploits both visual and textual information via a multimodal contrastive objective. Through experiments on a variety of semantic textual similarity tasks, we demonstrate that our approach consistently improves the performance across various datasets and pre-trained encoders. In particular, combining a small amount of multimodal data with a large text-only corpus, we improve the state-of-the-art average Spearman's correlation by 1.7%. By analyzing the properties of the textual embedding space, we show that our model excels in aligning semantically similar sentences, providing an explanation for its improved performance.

What makes a video task uniquely suited for videos, beyond what can be understood from a single image? Building on recent progress in self-supervised image-language models, we revisit this question in the context of video and language tasks. We propose the atemporal probe (ATP), a new model for video-language analysis which provides a stronger bound on the baseline accuracy of multimodal models constrained by image-level understanding. By applying this model to standard discriminative video and language tasks, such as video question answering and text-to-video retrieval, we characterize the limitations and potential of current video-language benchmarks. We find that understanding of event temporality is often not necessary to achieve strong or state-of-the-art performance, even compared with recent large-scale video-language models and in contexts intended to benchmark deeper video-level understanding. We also demonstrate how ATP can improve both video-language dataset and model design. We describe a technique for leveraging ATP to better disentangle dataset subsets with a higher concentration of temporally challenging data, improving benchmarking efficacy for causal and temporal understanding. Further, we show that effectively integrating ATP into full video-level temporal models can improve efficiency and state-of-the-art accuracy.

Despite progress in video-language modeling, the computational challenge of interpreting long-form videos in response to task-specific linguistic queries persists, largely due to the complexity of high-dimensional video data and the misalignment between language and visual cues over space and time. To tackle this issue, we introduce a novel approach called Language-guided Spatial-Temporal Prompt Learning (LSTP). This approach features two key components: a Temporal Prompt Sampler (TPS) with optical flow prior that leverages temporal information to efficiently extract relevant video content, and a Spatial Prompt Solver (SPS) that adeptly captures the intricate spatial relationships between visual and textual elements. By harmonizing TPS and SPS with a cohesive training strategy, our framework significantly enhances computational efficiency, temporal understanding, and spatial-temporal alignment. Empirical evaluations across two challenging tasks--video question answering and temporal question grounding in videos--using a variety of video-language pretrainings (VLPs) and large language models (LLMs) demonstrate the superior performance, speed, and versatility of our proposed LSTP paradigm.

Large language models have demonstrated impressive universal capabilities across a wide range of open-ended tasks and have extended their utility to encompass multimodal conversations. However, existing methods encounter challenges in effectively handling both image and video understanding, particularly with limited visual tokens. In this work, we introduce Chat-UniVi, a unified vision-language model capable of comprehending and engaging in conversations involving images and videos through a unified visual representation. Specifically, we employ a set of dynamic visual tokens to uniformly represent images and videos. This representation framework empowers the model to efficiently utilize a limited number of visual tokens to simultaneously capture the spatial details necessary for images and the comprehensive temporal relationship required for videos. Moreover, we leverage a multi-scale representation, enabling the model to perceive both high-level semantic concepts and low-level visual details. Notably, Chat-UniVi is trained on a mixed dataset containing both images and videos, allowing direct application to tasks involving both mediums without requiring any modifications. Extensive experimental results demonstrate that Chat-UniVi, as a unified model, consistently outperforms even existing methods exclusively designed for either images or videos.

Vision-language models advance multimodal representation learning by acquiring transferable semantic embeddings, thereby substantially enhancing performance across a range of vision-language tasks, including cross-modal retrieval, clustering, and classification. An effective embedding is expected to comprehensively preserve the semantic content of the input while simultaneously emphasizing features that are discriminative for downstream tasks. Recent approaches demonstrate that VLMs can be adapted into competitive embedding models via large-scale contrastive learning, enabling the simultaneous optimization of two complementary objectives. We argue that the two aforementioned objectives can be decoupled: a comprehensive understanding of the input facilitates the embedding model in achieving superior performance in downstream tasks via contrastive learning. In this paper, we propose CoMa, a compressed pre-training phase, which serves as a warm-up stage for contrastive learning. Experiments demonstrate that with only a small amount of pre-training data, we can transform a VLM into a competitive embedding model. CoMa achieves new state-of-the-art results among VLMs of comparable size on the MMEB, realizing optimization in both efficiency and effectiveness.

The recent success of the CLIP model has shown its potential to be applied to a wide range of vision and language tasks. However this only establishes embedding space relationship of language to images, not to the video domain. In this paper, we propose a novel approach to map video embedding space to natural langugage. We propose a two-stage approach that first extracts visual features from each frame of a video using a pre-trained CNN, and then uses the CLIP model to encode the visual features for the video domain, along with the corresponding text descriptions. We evaluate our method on two benchmark datasets, UCF101 and HMDB51, and achieve state-of-the-art performance on both tasks.

We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, pro-cessing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.

Forecasting what will happen next is a critical skill for general-purpose systems that plan or act in the world at different levels of abstraction. In this paper, we identify a strong correlation between a vision model's perceptual ability and its generalist forecasting performance over short time horizons. This trend holds across a diverse set of pretrained models-including those trained generatively-and across multiple levels of abstraction, from raw pixels to depth, point tracks, and object motion. The result is made possible by a novel generalist forecasting framework that operates on any frozen vision backbone: we train latent diffusion models to forecast future features in the frozen representation space, which are then decoded via lightweight, task-specific readouts. To enable consistent evaluation across tasks, we introduce distributional metrics that compare distributional properties directly in the space of downstream tasks and apply this framework to nine models and four tasks. Our results highlight the value of bridging representation learning and generative modeling for temporally grounded video understanding.

Learning discriminative spatiotemporal representation is the key problem of video understanding. Recently, Vision Transformers (ViTs) have shown their power in learning long-term video dependency with self-attention. Unfortunately, they exhibit limitations in tackling local video redundancy, due to the blind global comparison among tokens. UniFormer has successfully alleviated this issue, by unifying convolution and self-attention as a relation aggregator in the transformer format. However, this model has to require a tiresome and complicated image-pretraining phrase, before being finetuned on videos. This blocks its wide usage in practice. On the contrary, open-sourced ViTs are readily available and well-pretrained with rich image supervision. Based on these observations, we propose a generic paradigm to build a powerful family of video networks, by arming the pretrained ViTs with efficient UniFormer designs. We call this family UniFormerV2, since it inherits the concise style of the UniFormer block. But it contains brand-new local and global relation aggregators, which allow for preferable accuracy-computation balance by seamlessly integrating advantages from both ViTs and UniFormer. Without any bells and whistles, our UniFormerV2 gets the state-of-the-art recognition performance on 8 popular video benchmarks, including scene-related Kinetics-400/600/700 and Moments in Time, temporal-related Something-Something V1/V2, untrimmed ActivityNet and HACS. In particular, it is the first model to achieve 90% top-1 accuracy on Kinetics-400, to our best knowledge. Code will be available at https://github.com/OpenGVLab/UniFormerV2.

We investigate compositional structures in data embeddings from pre-trained vision-language models (VLMs). Traditionally, compositionality has been associated with algebraic operations on embeddings of words from a pre-existing vocabulary. In contrast, we seek to approximate representations from an encoder as combinations of a smaller set of vectors in the embedding space. These vectors can be seen as "ideal words" for generating concepts directly within the embedding space of the model. We first present a framework for understanding compositional structures from a geometric perspective. We then explain what these compositional structures entail probabilistically in the case of VLM embeddings, providing intuitions for why they arise in practice. Finally, we empirically explore these structures in CLIP's embeddings and we evaluate their usefulness for solving different vision-language tasks such as classification, debiasing, and retrieval. Our results show that simple linear algebraic operations on embedding vectors can be used as compositional and interpretable methods for regulating the behavior of VLMs.