Reproduce “Tracking People by Predicting 3D Appearance, Location & Pose” 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
Jathushan Rajasegaran, Georgios Pavlakos, Angjoo Kanazawa, Jitendra Malik
UC Berkeley researchers developed PHALP (Predicting Human Appearance, Location & Pose for Tracking), a multi-person tracking system that leverages 3D representations of people from monocular video. This approach improves tracking robustness in complex scenarios, reducing identity switches by enabling better handling of occlusions and viewpoint changes.
Key insights
- Tracking people in 3D space allows for robust handling of occlusions and viewpoint changes by leveraging object permanence and canonical appearance representations.
- Predicting future 3D states (appearance, location, pose) using learned models (linear regression for location, transformer for pose) significantly improves association accuracy.
- A probabilistic association framework that combines distances from independent 3D attributes (appearance, location, pose) into a combined cost function, with empirically optimized parameters, leads to superior tracking performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.