Reproduce “HOIGS: Human-Object Interaction 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

Taewoo Kim, Suwoong Yeom, Jaehyun Pyun, Geonho Cha, Dongyoon Wee, Joonsik Nam, Yun-Seong Jeong, Kyeongbo Kong and 1 more

The HOIGS framework reconstructs dynamic human-object interactions from monocular video by disentangling human and object motions with entity-specific deformation models and unifying them via a cross-attention interaction module. This approach achieved high visual quality and geometric accuracy on HOSNeRF, BEHAVE, and ARCTIC datasets, effectively reducing common reconstruction artifacts.

Key insights

  • Disentangling human and object representations and applying specialized deformation strategies for each entity (articulated human vs. rigid object) significantly improves the accuracy of complex HOI reconstruction.
  • Explicitly modeling bidirectional human-object dependencies through a cross-attention mechanism effectively mitigates interaction artifacts and ensures more physically plausible motion, even in close contact or occlusion.
  • Leveraging robust object initialization with diffusion priors and incorporating depth supervision enhances geometric fidelity, particularly for plausibly completing 3D geometry in occluded object regions.

Summary by alphaXiv.

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