360DVO Dataset

This is the official dataset from the paper 360DVO: Deep Visual Odometry for Monocular 360-Degree Camera, which is published on IEEE Robotics and Automation Letters (RA-L) 2026.

Project Page | Paper | Code

360DVO is a deep learning-based omnidirectional visual odometry (OVO) framework. It introduces a distortion-aware spherical feature extractor (DAS-Feat) and an omnidirectional differentiable bundle adjustment (ODBA) module. This repository provides the real-world OVO benchmark introduced in the study to facilitate evaluation in realistic settings.

Updates

• Jan. 9th, 2026: we update the names of each sequence for better understanding.
• Jan. 6th, 2026: we add project page and arxiv paper link.
• Jan. 4th, 2026: we release the 360DVO dataset with ground truth trajectories.

Details

360DVO Dataset is a large-scale real-world OVO benchmark emphasizing practical challenges across diverse environments and motions. It includes 20 sequences (~1k frames each) and all images are standardized to 3840x1920 at 10 FPS. Pseudo ground truth is reconstructed via SfM software Agisoft Metashape.

To facilitate rigorous evaluation, we partition the dataset into Easy and Hard subsets (10 each) by trajectory complexity and scene dynamics, from linear, static cases to aggressive rotations, lighting shifts, and dynamic occlusions.

Sequence	00	01	02	03	04	05	06	07	08	09
Easy	bridge night	canyon line	city driving	downhill biking	hongkong central	hongkong wanchai	mountains	shanghai driving	snowmobile	tokyo citywalk
Hard	canyon loop	dragon boat	drone racetrack	field	grove	indoor RC car	london bridge	ridge to lake	snowy mountain road	wingsuit

• Usage

# Authentication
from huggingface_hub import login
login()

# Download
from datasets import load_dataset
dataset = load_dataset("chris1004336379/360DVO")

We also provide a script eval.py for evaluation.

• Citation Information

@misc
{
  guo2026360dvodeepvisualodometry,
  title={360DVO: Deep Visual Odometry for Monocular 360-Degree Camera}, 
  author={Xiaopeng Guo and Yinzhe Xu and Huajian Huang and Sai-Kit Yeung},
  year={2026},
  eprint={2601.02309},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2601.02309},
}