TrackEverything: 3D de-duplication breaks long-horizon dense point tracking trade-off

zhenjun_zhao · x · 2026-09-29

Researchers introduced TrackEverything, a 3D point tracker that represents videos as persistent 3D scene tracks in world coordinates, decoupling model complexity from video duration. Three innovations: voxelization-based de-duplication at sliding-window boundaries to merge co-located tracks; decomposing tracking into an endpoint refiner plus a lightweight trajectory refiner for dynamic points; and 3D WAFT, replacing 4D correlation volumes with efficient feature sampling in the scene cloud. The authors claim it is the first 3D tracker able to track all visible points in videos over an hour long.

Related event: TrackEverything Enables Dense Long-Horizon Point Tracking in 3D Scenes(2 posts)→

Original post →

More from Embodied

Embodied channel →