Reproduce “Physically Plausible Human-Object Rendering from Sparse Views via 3D Gaussian Splatting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Weiquan Wang, Jun Xiao, Yi Yang, Yueting Zhuang, Long Chen

A framework named HOGS enables efficient and physically plausible rendering of human-object interactions from sparse camera views by integrating 3D Gaussian Splatting with explicit contact and separation losses. The approach achieves a PSNR of 30.68 dB on the HODome dataset, surpassing prior 3DGS methods by 3.4 dB, and renders at 162 FPS.

Key insights

  • Decoupling the learning dynamics of human and object Gaussians, while rasterizing them together, is essential for handling their distinct properties (deformable vs. rigid) and resolving complex occlusion relationships effectively.
  • Pre-trained, generalizable modules that refine sparse-view human pose and predict human-object contact regions provide crucial geometric priors, significantly improving robustness and computational efficiency of per-scene Gaussian optimization under limited visual input.
  • Explicit 3D contact and separation losses, when efficiently applied to predicted interaction zones via Signed Distance Fields (SDFs), are vital for enforcing physical plausibility and preventing geometric artifacts that 2D image supervision alone cannot resolve.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.