Reproduce “SmoothNet: A Plug-and-Play Network for Refining Human Poses in Videos” 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
Ailing Zeng, Lei Yang, Xuan Ju, Jiefeng Li, Jianyi Wang, Qiang Xu
SmoothNet, from researchers at The Chinese University of Hong Kong and Sensetime Group Ltd., is a plug-and-play temporal-only network designed to refine human pose estimations from videos. It demonstrably reduces temporal jitters, for example, achieving an 86.88% Accel reduction and an 8.82% MPJPE improvement on the AIST++ dataset using VIBE, and shows strong transferability across various pose estimators and modalities.
Key insights
- Focusing exclusively on temporal relations for each joint independently, rather than complex spatio-temporal correlations, effectively mitigates unbalanced and long-term jitters without introducing noisy inter-joint dependencies.
- Explicitly modeling first-order (velocity) and second-order (acceleration) motion information, as in the motion-aware design, improves both the fitting capability and overall accuracy of the smoothed poses.
- Fully-Connected Networks with global receptive fields are highly effective for learning motion continuity and correcting long-term jitters, often outperforming TCNs and Transformers for this specific temporal refinement task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.