Reproduce “Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, Kris Kitani
The paper introduces Observation-Centric SORT (OC-SORT), an enhanced multi-object tracking method that improves robustness to occlusion and non-linear motion while retaining the efficiency of the original SORT algorithm. This approach, which strategically leverages modern object detections, achieves state-of-the-art performance on various benchmarks, including MOT17, MOT20, and DanceTrack, running at over 700 FPS on a single CPU.
Key insights
- Simple, online, and real-time (SORT) paradigms can achieve state-of-the-art performance by intelligently leveraging the high accuracy of modern object detectors.
- Re-calibrating Kalman filter states with observation-anchored virtual trajectories effectively addresses temporal error magnification during occlusions.
- Incorporating motion direction consistency derived from observations into the association process enhances tracking robustness for non-linear motions, even with simple Kalman filters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.