AI
TrackEverything: Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 23/09/2026, 20:00:00 UTC
Briefing created: 28/09/2026, 05:26:55 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/09/2026, 05:18:24 UTC.
Briefings
TrackEverything is a 3D point tracker that resolves the trade-off between long-horizon sparse tracking and short-horizon dense tracking by representing videos as persistent 3D scene tracks in world coordinates. The system decouples model complexity from video duration, scaling with unique physical scene geometry. It introduces three innovations: voxelization-based de-duplication at sliding-window boundaries to merge co-located tracks, an endpoint refiner predicting destination and dynamic classification, and a lightweight trajectory refiner for dynamic points. To the best of the authors' knowledge, it is the first 3D tracker capable of tracking all visible points across videos exceeding 1000 frames within 40 GB of GPU memory. On TAPVid-3D, it outperforms all open-source all-frame dense 3D trackers by more than 20% APD on short clips while remaining competitive with state-of-the-art sparse trackers on long sequences.
Key points
- TrackEverything represents videos as persistent 3D scene tracks in world coordinates to decouple model complexity from video duration.
- The system uses voxelization-based de-duplication at sliding-window boundaries to merge co-located tracks and prevent redundant accumulation.
Context and limitations
The study is limited to the TAPVid-3D dataset and does not report results on other datasets or longer video sequences beyond the 1000-frame threshold.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 28/09/2026, 05:26:55 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidjej464q6djfo2kgpbhfyehzpkmjst3blogsz6b6gvdhl2xda4gy