Researchers introduced TrackEverything, a 3D point tracker for long videos. It represents videos as persistent 3D scene tracks in world coordinates. The model uses voxelization to merge co-located tracks at sliding-window boundaries. It splits tracking into endpoint refinement and static-versus-dynamic classification. A lightweight trajectory refiner decodes dense trajectories for dynamic points only. It replaces 4D correlation volumes with 3D WAFT for efficient feature sampling. TrackEverything is the first 3D tracker to track all visible points across 1000+ frames within 40 GB of GPU memory. On TAPVid-3D, it outperforms open-source dense 3D trackers by over 20% APD on short clips. It remains competitive with state-of-the-art sparse trackers on long sequences. The paper was submitted to arXiv on 24 September 2026.

• 20% APD improvement on TAPVid-3D
• 1000+ frames tracked
• 40 GB GPU memory limit