Reproduce “AnthroTAP: Learning Point Tracking with Real-World Motion” 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
Inès Hyeonsu Kim, Seokju Cho, Jahyeok Koo, Junghyun Park, Jiahui Huang, Honglak Lee, Joon-Young Lee, Seungryong Kim
AnthroTAP generates point-track labels from 3D human-body meshes by projecting mesh vertices into video, estimating visibility with ray-casting, and filtering tracks against optical flow. They train LocoTrack on 1.4K Anthro-LD videos and Kubric data. On TAP-Vid-DAVIS First at 256 × 256, Anthro-LocoTrack scores 64.8 AJ versus 63.0 for the Kubric-trained baseline, with labels limited by mesh and visibility errors.
Key insights
- Real human motion, particularly from dance videos, provides a rich source of complex and diverse motion patterns (non-rigid deformation, articulated motion, frequent occlusions) highly valuable for training robust point trackers.
- Distilling knowledge from sophisticated 3D human mesh recovery models offers a strong, structured supervision signal for 2D point tracking, proving more effective than weaker self-supervision or reliance on synthetic data for generalization.
- The inherent visual diversity and appearance characteristics of real-world data are more crucial for learning generalizable tracking capabilities than achieving perfect ground truth precision in synthetic environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.