README.md

metadata

license: other
license_name: source-dataset-terms
viewer: false
extra_gated_heading: Acknowledge license to access the repository
extra_gated_prompt: >-
  This dataset is derived from HO-3D, iTACO and YCBInEOAT. By accessing it, you
  agree to the terms of the original datasets:
  [HO-3D](https://github.com/shreyashampali/ho3d) (academic, non-commercial
  research only),
  [iTACO](https://huggingface.co/datasets/3dlg-hcvc/video2articulation) (CC
  BY-NC 4.0) and the [PartNet-Mobility terms of
  use](https://sapien.ucsd.edu/about#term) it builds on, and
  [YCBInEOAT](https://github.com/wenbowen123/iros20-6d-pose-tracking). Use is
  limited to non-commercial research and education; do not redistribute the
  data, and cite the source datasets.
extra_gated_fields:
  Name: text
  Affiliation: text
  Purpose: text
  I agree to use this dataset for non-commercial research and education only, under the terms of the source datasets: checkbox
extra_gated_button_content: Agree and access repository

MoSE3 SE(3) evaluation caches

Pre-built benchmark caches for the SE(3) evaluation of MoSE3: Learning World-Space SE(3) at Every Pixel (NeurIPS 2026). The evaluation code is on the evaluation branch of mose3-tracker/MoSE3.

hf download mose3-tracker/mose3-eval --repo-type dataset --local-dir data/se3_bench

ho3d/        54 clips
itaco/       55 clips
ycbineoat/    9 clips

Each clip is one <clip>.npz with T = 50 frames (49 on YCBInEOAT) at 480 x 640, N = 3072 query pixels at frame 0, and the GT 3D tracks, visibility and per-query object poses. See the evaluation README for the keys.

Source datasets

The clips are derived from the following datasets. Their original terms apply (non-commercial research use); please cite them if you use these caches.

• HO-3D: Hampali et al., HOnnotate: A Method for 3D Annotation of Hand and Object Poses, CVPR 2020. arXiv:1907.01481
• iTACO: Peng et al., iTACO: Interactable Digital Twins of Articulated Objects from Casually Captured RGBD Videos. arXiv:2506.08334, built on PartNet-Mobility.
• YCBInEOAT: Wen et al., se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains, IROS 2020. arXiv:2007.13866

Citation

@inproceedings{cheng2026mose3,
  title     = {{MoSE3}: Learning World-Space {SE(3)} at Every Pixel},
  author    = {Cheng, Jiahuan and Li, Zhiyi and Xia, Tian and
               Cai, Ruojin and Du, Yilun and Wang, Qianqian},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}