Reproduce “ProTracker: Probabilistic Integration for Robust and Accurate Point Tracking” on OpenResearch

About this paper

Tingyang Zhang, Chen Wang, Zhiyang Dou, Qingzhe Gao, Jiahui Lei, Baoquan Chen, Lingjie Liu

ProTracker introduces a probabilistic integration framework that combines local optical flow predictions with global long-term correspondences to achieve robust and accurate long-term dense point tracking. This approach consistently outperforms previous optimization-based methods and demonstrates superior positional accuracy compared to many supervised trackers on TAP-Vid and BADJA benchmarks, while effectively mitigating drift and handling occlusions.

Key insights

• Probabilistic integration effectively combines local optical flow for fine-grained prediction with global semantic information from long-term correspondences, balancing precision and robustness.
• Hybrid filtering, composed of object-level and geometry-aware feature filters, significantly enhances robustness by preemptively removing unreliable predictions and enforcing global context and semantic consistency.
• Bidirectional integration, coupled with the joint use of optical flow and long-term key points, is crucial for mitigating accumulated drift and successfully re-localizing points after extended occlusions and disappearances.

Summary by alphaXiv.