File size: 3,147 Bytes
6dc2e65
 
 
 
 
 
 
 
 
 
 
7478ea2
6dc2e65
 
 
 
 
 
 
6c8057a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 6174743543.802
    num_examples: 1142
  download_size: 6408762030
  dataset_size: 6174743543.802
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---


# GeoBench (GeoVista Bench)

GeoBench is a collection of real-world panoramas with rich metadata for evaluating geolocation models. Each sample corresponds to one panorama identified by its `uid` and includes both the original high-resolution imagery and a lightweight preview for rapid inspection.

## Dataset Structure

- `id`: unique identifier (same as `uid` from the original data).
- `raw_image_path`: relative path (within this repo) to the source panorama under `raw_image/<uid>/`.
- `preview`: compressed JPEG preview (<=1M pixels) under `preview_image/<uid>/`. This is used by HF Dataset Viewer.
- `metadata`: JSON object storing capture timestamp, location, pano_id, city, and other attributes. Downstream users can parse it to obtain lat/lng, city names, multi-level location tags, etc.
- `data_type`: string describing the imagery type. If absent in metadata it defaults to `panorama`.

All samples are stored in a Hugging Face-compatible parquet file at `data/<split>/data-00000-of-00001.parquet`, with additional metadata in `dataset_info.json`.

## Working with GeoBench

1. Clone/download this folder (or pull it via `huggingface_hub`).
2. Load the parquet file using Python:
   ```python
   from datasets import load_dataset

   ds = load_dataset('path/to/this/folder', split='train')
   sample = ds[0]
   ``
   `sample["preview"]` loads directly as a PIL image; `sample["raw_image_path"]` points to the higher-quality file if needed.
3. Use the metadata to drive evaluation logic, e.g., compute city-level accuracy, filter by `data_type`, or inspect specific regions.

## Notes

- Raw panoramas retain original filenames to preserve provenance.
- Preview images are resized to reduce storage costs while remaining representative of the scene.
- Ensure you comply with the dataset’s license (`dataset_info.json`) when sharing or modifying derived works.

## Related Resources

	•	GeoVista model (RL-trained agentic VLM used in the paper):
https://huggingface.co/LibraTree/GeoVista
	•	GeoVista-Bench (previewable variant):
A companion dataset with resized JPEG previews intended to make image preview easier in the Hugging Face dataset viewer:
https://huggingface.co/datasets/LibraTree/GeoVistaBench
(Same underlying benchmark; different packaging / image formats.)
	•	Paper page on Hugging Face:
https://huggingface.co/papers/2511.15705


## Citation
```
@misc{wang2025geovistawebaugmentedagenticvisual,
      title        = {GeoVista: Web-Augmented Agentic Visual Reasoning for Geolocalization},
      author       = {Yikun Wang and Zuyan Liu and Ziyi Wang and Pengfei Liu and Han Hu and Yongming Rao},
      year         = {2025},
      eprint       = {2511.15705},
      archivePrefix= {arXiv},
      primaryClass = {cs.CV},
      url          = {https://arxiv.org/abs/2511.15705},
}
```