Reproduce “ReCHOIR: Contact-guided Human Object Interaction Retargeting to Diverse Characters” 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
Chaelin Kim, Seokhyeon Hong, Kwan Yun, Soojin Choi, Inseo Jang, Junyong Noh
ReCHOIR retargets human-object interactions across characters with different skeletons. It encodes motion as whole-body and body-part tokens, then uses source-object and contact cues to adjust target motion and predict object translation while retaining source orientation. On OMOMO-test, it reaches 0.9713 contact F1 at 0.421 ms/frame; PAME + Optim scores 0.9891 at 199.611 ms/frame, showing a speed–contact tradeoff.
Key insights
- Decomposing HOI retargeting into a general human motion prior followed by contact-guided refinement is effective, particularly in the absence of paired HOI supervision data.
- A part-aware latent representation, as learned by PAME, enables better generalization and fine-grained adaptation of human motion across diverse skeletal structures.
- A ControlNet-style residual conditioning mechanism effectively injects interaction cues into a pretrained motion decoder, guiding and refining motion based on interaction constraints without discarding the learned human motion prior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.