Projects Last Projects
Real-Time Long-Range Point Tracking Using Image Foundation Models
Dor Danino & Hanoch Califa
Supervised by Asaf Singer
We present a real-time dense point tracking framework that builds upon the DINO-Tracker baseline, aiming to improve point tracking performance by addressing the limitations of standard DINOv2 features. Our approach integrates optimized DINOv2-based model - DINO FeatUp - which enhance the spatial quality of DINOv2 embeddings. These adapted features enable more accurate and stable point tracking across frames. Unlike the original DINO-Tracker, which relies on test-time training, our pipeline is optimized for real-time operation. our method consistently outperforms the use of raw DINOv2 features within the same tracking flow, highlighting the benefits of task-specific DINOv2 adaptation.
Please, see project papers.
Please, see project report.
Please, see final presentation.