haodongli/DA-2

DA²: Depth Anything in Any Direction

DA² predicts dense, scale-invariant distance from a single 360° panorama in an end-to-end manner, with remarkable geometric fidelity and strong zero-shot generalization.

⬇️ Download

1. Download the datasets (please see here for the environment setup):

cd [YOUR_DATA_DIR]
huggingface-cli login
hf download --repo-type dataset haodongli/DA-2 --local-dir [YOUR_DATA_DIR]

2. Merge parts into one *.tar.gz file:

   DATASET_NAME in [hypersim_pano, vkitti_pano, mvs_synth_pano, unreal4k_pano, 3d-ken-burns_pano, dynamic_replica_v2_pano]

cat [DATASET_NAME]/part_* > [DATASET_NAME].tar.gz

3. Check the MD5:

md5sum -c [DATASET_NAME]_checksum.md5

4. If correct, then we can unzip it:

tar -zxvf [DATASET_NAME].tar.gz

5. The data samples will be exported in [DATASET_NAME]/.

🎮 Usage

1. The dietance values from the pixel to the 360° camera is stored in depth.png. I also provided depth_vis.png just for visualization.
2. Please refer the code below to load the depth values from depth.png:

depth = cv2.imread('path/to/depth.png', cv2.IMREAD_UNCHANGED)
depth = depth.astype(np.float32)
depth = depth[:,:,0]
depth = depth * SCALE
depth = torch.from_numpy(depth)

3. Please see the below table for the SCALE of different curated dataset:

   Curated dataset	Scale
   Hypersim	40.0 / 65535.0
   VKITTI, MVS-Synth, 3D-Ken-Burns	1.0 / 256.0
   UnrealStereo4K	80.0 / 65535.0
   DynamicReplica	20.0 / 65535.0

4. The valid masks of the depth maps can be obtained via:

valid_mask = torch.logical_and(
    (depth > 1e-5), (depth < 80.0)
).bool()

🎓 Citation

If you find these datasets useful, please consider citing 🌹:
Please note there is a blankspace (" ") between "DA" and "$^{2}$" in the BibTeX of our paper 🥹

@article{li2025depth,
  title={DA $^{2}$: Depth Anything in Any Direction},
  author={Li, Haodong and Zheng, Wangguangdong and He, Jing and Liu, Yuhao and Lin, Xin and Yang, Xin and Chen, Ying-Cong and Guo, Chunchao},
  journal={arXiv preprint arXiv:2509.26618},
  year={2025}
}

@inproceedings{roberts2021hypersim,
  title={Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding},
  author={Roberts, Mike and Ramapuram, Jason and Ranjan, Anurag and Kumar, Atulit and Bautista, Miguel Angel and Paczan, Nathan and Webb, Russ and Susskind, Joshua M},
  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
  pages={10912--10922},
  year={2021}
}

@article{cabon2020virtual,
  title={Virtual kitti 2},
  author={Cabon, Yohann and Murray, Naila and Humenberger, Martin},
  journal={arXiv preprint arXiv:2001.10773},
  year={2020}
}

@inproceedings{huang2018deepmvs,
  title={Deepmvs: Learning multi-view stereopsis},
  author={Huang, Po-Han and Matzen, Kevin and Kopf, Johannes and Ahuja, Narendra and Huang, Jia-Bin},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={2821--2830},
  year={2018}
}

@inproceedings{tosi2021smd,
  title={Smd-nets: Stereo mixture density networks},
  author={Tosi, Fabio and Liao, Yiyi and Schmitt, Carolin and Geiger, Andreas},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={8942--8952},
  year={2021}
}

@article{niklaus20193d,
  title={3d ken burns effect from a single image},
  author={Niklaus, Simon and Mai, Long and Yang, Jimei and Liu, Feng},
  journal={ACM Transactions on Graphics (ToG)},
  volume={38},
  number={6},
  pages={1--15},
  year={2019},
  publisher={ACM New York, NY, USA}
}

@inproceedings{karaev2023dynamicstereo,
  title={Dynamicstereo: Consistent dynamic depth from stereo videos},
  author={Karaev, Nikita and Rocco, Ignacio and Graham, Benjamin and Neverova, Natalia and Vedaldi, Andrea and Rupprecht, Christian},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13229--13239},
  year={2023}
}