ReLift takes the 2D tracks and depth from existing tools and repairs them into dense 3D tracks from monocular video.
Approach: A feed-forward trajectory corrector alternates attention along and across trajectories to fix corrupted positions and fill in occluded motion.
Training: It trains only on 3D trajectories with simulated lifting errors, with no image-based training, so one model works across trackers and geometry providers without retraining.
Results: On WorldTrack it reports state-of-the-art mean accuracy, 2.1 APD ahead of the strongest baselines on all points and 3.0 on dynamic points.
Speed: Faster configurations stay competitive with leading 3D trackers and run faster than 4RC.
Sign in to comment.
TL;DR of the paper, generated automatically.
ReLift takes the 2D tracks and depth from existing tools and repairs them into dense 3D tracks from monocular video.