GenPT's generative capabilities can be leveraged to improve point trajectory estimates by utilizing a best-first search on generated samples during inference, guided by the model’s own confidence of its predictions.

GenPT is trained with a modified flow matching setup, incorporating iterative refinements, a window-dependent prior for cross-window consistency, and a variance schedule tuned for point coordinates.

It is a generative framework for modelling multi-modal point trajectories.

Tracking the location of a point through a video is an ill-posed problem, where multiple plausible solutions can arise during visual obfuscations, such as occlusion.

Current point trackers are discriminative and fail to model the multi-modality inherent in point tracking.

Generative Point Tracking with Flow Matching

My latest project with Adam W. Harley, @csprofkgd.bsky.social, Derek Nowrouzezahrai, @chrisjpal.bsky.social.

Project page: mtesfaldet.net/genpt_projpa...
Paper: arxiv.org/abs/2510.20951
Code: github.com/tesfaldet/ge...