CoTracker3

CoTracker3 is a unified framework for object and point tracking that operates across video, images, and 3D scenes reconstructed from multi-view data. The tool is designed to track both sparse and dense points over time, enabling consistent correspondence of pixels or features even under large motion, occlusions, and complex deformations. CoTracker3 supports long-range tracking by integrating temporal information and geometric cues, making it suitable for challenging real-world sequences rather than just short clips or synthetic benchmarks.

A key capability is its ability to jointly reason about appearance and 3D structure, allowing more stable tracking on dynamic scenes and camera motions. The system can be applied to tasks such as motion analysis, video editing, visual effects, 3D reconstruction refinement, and robotics perception.