Technion – Israel Institute of Technology and NVIDIA (Assaf Singer, Maroun Khriesh, Noam Rotstein, Yoni Kasten, Or Litany) · 3D trajectory repair

ReLift: Dense 3D Tracking as Lifted Trajectory Repair

ReLift is a research paper from Technion and NVIDIA on dense 3D tracking from monocular video. It takes 2D tracks and depth estimates from pretrained models and lifts them into 3D. A feed-forward trajectory corrector then repairs lifting errors and completes motion through occlusion, alternating attention along and across trajectories. It is trained only on 3D trajectories with simulated lifting errors, so it uses geometry alone and works across different trackers and geometry models without retraining. On WorldTrack it reports state-of-the-art mean accuracy, improving on the strongest baselines by 2.1 APD on all points and 3.0 on dynamic points. Faster configurations stay competitive with leading 3D trackers while running 2.7-4.4× faster than 4RC. The page gives an interactive demo and an arXiv reference (2610.42136), but no licence, code release or version is stated in the text.