3D Point Tracking

3D point tracking follows points on objects or across whole scenes through video, recovering their motion in three dimensions rather than only in image coordinates. Methods combine visual correspondence with depth, multi-view geometry, point-cloud motion, and temporal constraints; newer approaches also adapt video foundation models, diffusion transformers, and state-space models for feed-forward or long-range tracking. The task supports dynamic-scene reconstruction, motion analysis, and robot learning, and is evaluated on benchmarks such as TAPVid-3D, PointOdyssey, KITTI, and nuScenes. A central challenge is maintaining accurate geometry and point identity through occlusion and long sequences without making inference too slow or memory-intensive.