TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
TrackEverything claims dense 3D point tracking past 1,000 video frames within 40 GB of GPU memory.
The paper frames the advance as escaping the usual split between long-range sparse tracking and short-clip dense tracking. Its method stores persistent 3D scene tracks in world coordinates, then de-duplicates repeated surface observations at sliding-window boundaries. It refines endpoints first, separates static from dynamic points, and only decodes dense trajectories for the dynamic set. On TAPVid-3D, the authors report more than 20% APD gains over open-source all-frame dense 3D trackers on short clips while staying competitive with sparse trackers on long sequences. ArXiv · AI/CL/LG's note
The paper frames the advance as escaping the usual split between long-range sparse tracking and short-clip dense tracking. Its method stores persistent 3D scene tracks in world coordinates, then de-duplicates repeated surface observations at sliding-window boundaries. It refines endpoints first, separates static from dynamic points, and only decodes dense trajectories for the dynamic set. On TAPVid-3D, the authors report more than 20% APD gains over open-source all-frame dense 3D trackers on short clips while staying competitive with sparse trackers on long sequences. ArXiv · AI/CL/LG's note
score 6