ympradyumna/physic-eval-release

PhySIC evaluation release

Per-frame predictions of PhySIC (SIGGRAPH Asia 2025, arXiv) on PROX-quantitative (178 frames) and RICH (100 frames), with the reconstructed scene each one was optimised against. They reproduce the paper's Table 2 with the evaluation code in the repository's eval_release/, which downloads this data on first run.

What is here

release_data/prox/<recording>/<frame>.npz        compact tier, 15 MB
release_data/rich/<frame>.npz                                  9.8 MB
release_data_full/prox/<recording>/<frame>.npz   full scene clouds, 7.6 GB
release_data_full/rich/<frame>.npz                                  5.6 GB

Both tiers hold the same predictions and give identical numbers. The compact tier keeps the scene points the contact metric can reach (3.6k per PROX frame, ~4.5k per RICH frame); the full tier is the complete reconstructed cloud (2.0M / ~2.6M points, floor plane included).

Per-frame fields

field	what it is
init_*, final_*	SMPL-X parameters (betas + 6D rotations) and camera translation, before and after scene-aware optimisation
K	predicted pinhole intrinsics
contact_deco	the image-based contact prediction the optimiser was given
scene_points, scene_normals	reconstructed scene, metric scale, camera frame
visibility_mask	body vertices projecting into the segmented person (PROX only)
contact_pred_reference	contact mask as computed by the pipeline, re-derived and checked by the evaluation

No dataset imagery, depth maps, meshes or ground truth are included. Ground truth comes from your own downloads of PROX and RICH, and the bodies use SMPL-X, each under its own licence.

Licence

Released for non-commercial scientific research only, under the same terms as PROX and RICH.

Citation

@inproceedings{ym2025physic,
  author    = {Yalandur Muralidhar, Pradyumna and Xue, Yuxuan and Xie, Xianghui and Kostyrko, Margaret and Pons-Moll, Gerard},
  title     = {PhySIC: Physically Plausible 3D Human-Scene Interaction and Contact from a Single Image},
  journal   = {SIGGRAPH Asia 2025 Conference Papers},
  year      = {2025},
}