Reproduce “Scene-aware Human Motion Forecasting via Mutual Distance Prediction” 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
Chaoyue Xing, Wei Mao, Miaomiao Liu
This research introduces a "mutual distance" representation to comprehensively model human-scene interaction, enabling a two-step prediction pipeline that forecasts 3D human motions. The method yields more accurate and physically plausible predictions, consistently outperforming state-of-the-art baselines across multiple datasets by reducing artifacts like scene penetration and ghost motion.
Key insights
- Explicitly modeling both the local pose of the entire human body and its global translation relative to the 3D scene through a combined "mutual distance" representation is crucial for generating physically plausible motions.
- The combination of per-vertex signed distances and scene-independent per-basis point distances effectively resolves the ambiguity inherent in purely local interaction metrics, providing robust global positioning cues.
- Representing the 3D scene as an SDF volume offers consistency with the mutual distance concept and allows for efficient querying, which is vital for calculating consistency losses during training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.