Reproduce “HUMOF: Human Motion Forecasting in Interactive Social Scenes” 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
Caiyi Sun, Yujing Sun, Xiao Han, Zemin Yang, Jiawei Liu, Xinge Zhu, Siu Ming Yiu, Yuexin Ma
HUMOF forecasts a person’s movement and body pose from past motion, nearby scene context, and other people’s motion. It represents human–human and human–scene interactions at multiple scales, then injects broad features before fine ones. On HOI-M³, it records mean path and pose errors of 174.6 and 117.9 millimeters, versus 184.8 and 121.6 for the best listed competitors. Dynamic-object evaluation remains preliminary and data-limited.
Key insights
- Explicitly modeling both human-human and human-scene interactions hierarchically is crucial for accurate human motion forecasting in dynamic, interactive environments.
- A coarse-to-fine injection strategy for hierarchical interaction features, where high-level context informs early reasoning and fine-grained details refine later stages, optimizes feature utilization for enhanced prediction.
- An adaptive frequency-domain rescaling mechanism within a Transformer architecture effectively guides the model's focus from coarse motion trends to fine-grained details, improving the plausibility and accuracy of predictions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.