OC-SORT (CVPR 2023) is a PyTorch multi-object tracker that fixes Kalman filter drift in SORT by re-updating tracks from observations. It takes detections, with YOLOX weights provided, and can be combined with BYTE association and cost metrics such as GIoU and CIoU. Deployment paths cover ONNX, TensorRT, ncnn and C++.
Who it is for: Computer vision engineers needing online multi-object tracking
Strengths
Association alone runs at about 700 FPS on an i9 CPU when detections are given
Reported results on MOT17, MOT20, KITTI, DanceTrack and CroHD in the README
Deployment docs for ONNX, TensorRT and ncnn, plus a C++ implementation
Pluggable cost metrics (GIoU, CIoU) and BYTE integration
Weaknesses
Pure motion model; appearance matching needs the separate Deep OC-SORT project
Detection stage is tied to YOLOX codebase and weights
No tagged releases; install and usage steps live in separate docs files
Benchmarks use shared hyper-parameters, so per-dataset tuning is left to the user