🚀Introducing Protracker: Inspired by Kalman filter, we tackle point tracking with a robust probabilistic approach. 🌟Our method integrates multiple predictions from both optical flow and semantic correspondences in a unified framework with probabilistic fusion. This ensures to generate smooth and accurate trajectories. ProTracker achieves state-of-the-art performance among self-supervised methods across multiple benchmarks and enhanced robustness in challenging scenarios like occlusion, similar regions, and low-feature areas. 💡Protracker offers a probabilistic framework that combines information of different granularity and semantics, paving the way for new advancements in tracking any point. 🔍#PointTracking #ComputerVision Project page: michaelszj.github.io/protracker/ Paper: arxiv.org/abs/2501.03220 Code: michaelszj.github.io/protracker/
