TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
TrackEverything claims dense 3D tracking past 1,000 video frames within 40 GB of GPU memory.
The paper says it does this by storing persistent 3D scene tracks in world coordinates, so compute scales with unique scene geometry rather than video length. It de-duplicates co-located tracks at sliding-window boundaries, then refines endpoints and only decodes dense trajectories for dynamic points. The authors report more than 20% APD gains over open-source all-frame dense 3D trackers on TAPVid-3D short clips, while staying competitive with sparse trackers on long sequences. Source: HF Daily Papers' note.
The paper says it does this by storing persistent 3D scene tracks in world coordinates, so compute scales with unique scene geometry rather than video length. It de-duplicates co-located tracks at sliding-window boundaries, then refines endpoints and only decodes dense trajectories for dynamic points. The authors report more than 20% APD gains over open-source all-frame dense 3D trackers on TAPVid-3D short clips, while staying competitive with sparse trackers on long sequences. Source: HF Daily Papers' note.
score 6