OmniRob

OmniRob is a real-world omnidirectional dataset for panoramic novel-view synthesis and 3D reconstruction. It contains data captured by a Unitree Go2 quadruped robot with a panoramic annular camera and by an unmanned aerial vehicle (UAV) with a panoramic imaging system.

The dataset accompanies Spherical-GOF, accepted at IEEE/RSJ IROS 2026. Camera poses and colored point clouds were reconstructed with OpenSfM and exported to the OmniPhotos-compatible JSON format used by Spherical-GOF.

Dataset overview

Scene	Platform / image type	Images	Train / Test	Resolution
1	Go2, annular panorama	92	69 / 23	4096 x 512
2	Go2, annular panorama	81	61 / 20	4096 x 512
3	UAV, full equirectangular panorama	100	75 / 25	7680 x 3840
4	UAV, full equirectangular panorama	100	75 / 25	7680 x 3840
5	Cropped UAV panorama corresponding to Scene 3	100	75 / 25	7680 x 1281
6	Cropped UAV panorama corresponding to Scene 4	100	75 / 25	7680 x 1281

Scenes 3 and 5 share the same camera poses and point cloud. Scenes 4 and 6 do the same.

Directory structure

scene/
├── images/                  # panoramic RGB images
├── data_views.json          # image metadata and view/pose IDs
├── data_extrinsics.json     # camera rotations and centers
├── train.txt                # training image stems, one per line
├── test.txt                 # test image stems, one per line
└── pcd.ply                  # colored initialization point cloud

The JSON files use the OpenMVG-style schema used by OmniPhotos. Each view stores its filename, image dimensions, view ID, and pose ID. Each corresponding extrinsic stores a 3 x 3 rotation matrix and camera center. The PLY files store XYZ coordinates and RGB colors.

Use with Spherical-GOF

After installing Spherical-GOF, train a scene with:

python train.py \
    -s /path/to/OmniRob/1 \
    -m output/omnirob/1 \
    --eval \
    -r 1

--eval preserves the provided test split. Increase -r to downsample the input if a full-resolution scene does not fit in GPU memory.

License

OmniRob is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. You may share and adapt the dataset for non-commercial purposes with appropriate attribution. See LICENSE for the full terms.

Citation

If you use OmniRob, please cite Spherical-GOF:

@article{yang2026sphericalgof,
  title   = {Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction},
  author  = {Yang, Zhe and Zhao, Guoqiang and Wu, Sheng and Luo, Kai and Yang, Kailun},
  journal = {arXiv preprint arXiv:2603.08503},
  year    = {2026}
}