ReLift

ReLift is a research model for dense 3D point tracking from monocular video, described in the paper "ReLift: Dense 3D Tracking as Lifted Trajectory Repair" (Singer, Khriesh, Rotstein, Kasten, Litany; arXiv 2610.42136). It is listed under Nvidia as vendor. Rather than learning tracking, depth and occlusion handling jointly, it takes the output of existing 2D trackers and 3D geometry models, lifts the tracks into 3D, and corrects them.

The core component is a trajectory corrector that alternates attention along and across trajectories to predict 3D residuals. These residuals repair correspondence and depth errors and complete motion during occlusion, where lifted depth would otherwise come from the occluder. The corrector uses only geometric inputs and is trained on 3D trajectories with simulated lifting errors. This makes it independent of images and of any specific tracker or geometry model, so it works across them without retraining.

The authors report state-of-the-art mean accuracy on WorldTrack, with gains of 2.1 APD on all points and 3.0 APD on dynamic points over the strongest baselines. Faster configurations run 2.7–4.4× faster than 4RC. Demos use CoWTracker or AllTracker tracks with VGGT-Ω geometry. A YouTube creator reports that code, weights and training scripts were released, but no repository, Hugging Face page or license appeared in the provided material, so those details are unconfirmed.