Foundation models have rapidly become a central substrate for visual understanding and generation. They encode rich semantic and geometric priors that can be harnessed for a wide range of vision problems. This work studies how such priors can be combined with architectural inductive biases of deep neural networks to obtain optimization-based, lightweight yet powerful methods for two challenging tasks: (i) long-range video point tracking and, (ii) dynamic 3D reconstruction from monocular video.
The first part focuses on DINO-Tracker, a dense long-range point tracking method that leverages self-supervised DINO-ViT features as a semantic prior. To adapt these features for precise tracking, DINO-Tracker combines test-time optimization with a CNN-based feature refinement model, producing localized and temporally consistent features that support robust tracking across long occlusions. The second part introduces DRoPS, a method for dynamic 3D reconstruction from monocular video. DRoPS combines the power of pre-trained tracking and depth models with a novel Deep Motion Prior -- a CNN-based parameterization of the motion field, achieving state-of-the-art results in monocular dynamic 3D reconstruction.
Narek is a final-year PhD student (direct-track) at the Weizmann Institute of Science, advised by Prof. Tali Dekel. Previously, he was a Research Scientist Intern at Meta Reality Labs, advised by Jonathon Luiten. His research interests are in generative AI, 3D/4D reconstruction and tracking, and interpretability of vision foundation models for unveiling novel applications.