TrackEverything:
Long Horizon Dense Tracking via De-Duplicating 3D Scene Representations

Anonymous Authors

TrackEverything tracks all points across all frames of videos with 1000+ frames by representing the video as persistent 3D scene tracks in world coordinates.

In all visualizations, we show trails for dynamic points only. The method tracks all points and accounts for camera motion; we omit static-point trails in the 2D renderings for clarity.

Rebuttal

Baseline Comparisons

DELTA-v2

CoTracker3

SpatialTracker-v2

VDPM

Ours

OOM

Each row is one sequence: input video, 3D tracks, and 2D tracks. Scroll sideways for more; click any clip to enlarge.

Long Video Evaluation

Dense tracks on full-length sequences (1000+ frames). Click any clip to enlarge. We recommend clicking the dance videos to enjoy tracking with music. Drag or scroll the strip to see more.

More Qualitative Results

Dense 3D tracks rendered in the input view. Drag or scroll the strip to see the rest; click any clip to enlarge.

Static / Dynamic Classification

Per-window classification decides where trajectories are decoded. Red points are classified as dynamic, blue points as static. Drag or scroll the strip to see the rest; click any clip to enlarge.

Comparison with Baselines

Qualitative comparison against prior trackers on the same sequences. Notice how prior methods mostly track points from the first frame, while TrackEverything begins tracking a point as soon as it appears in the video. Each row is one sequence; click any clip to enlarge. Notice how prior methods mostly track points from the first frame, while TrackEverything begins tracking a point as soon as it appears in the video.

DELTA-v2

CoTracker3

SpatialTracker-v2

Ours

Failure Cases

Tracking degrades under fast, large-displacement motion and under repetitive or low-texture backgrounds, where the lifted 3D correspondences become ambiguous.

BibTeX

@misc{trackeverything2026,
  title     = {TrackEverything: Long Horizon Dense Tracking
               via De-Duplicating 3D Scene Representations},
  author    = {Anonymous Authors},
  year      = {2026}
}