ReLift: Dense 3D Tracking as Lifted Trajectory Repair
ReLift improves 3D tracking by repairing trajectories built from existing 2D trackers and depth models, without retraining for each combination.
The framework corrects position errors and fills gaps when tracked points disappear behind objects. It trains on 3D trajectories with simulated errors, using geometry alone rather than images. On WorldTrack, the authors report state-of-the-art mean accuracy, beating the strongest baselines by 2.1 APD on all points and 3.0 on dynamic points. Faster configurations remain competitive with leading 3D trackers while running faster than 4RC.
HF Daily Papers' note
The framework corrects position errors and fills gaps when tracked points disappear behind objects. It trains on 3D trajectories with simulated errors, using geometry alone rather than images. On WorldTrack, the authors report state-of-the-art mean accuracy, beating the strongest baselines by 2.1 APD on all points and 3.0 on dynamic points. Faster configurations remain competitive with leading 3D trackers while running faster than 4RC.
HF Daily Papers' note
score 5