ByteTrack

ByteTrack tackles multi-object tracking with a deliberately simple idea: instead of discarding low-confidence detection boxes, which often correspond to occluded objects, its BYTE association step matches them against existing tracklets to recover true objects and filter out background noise. That single change reduces missed detections and fragmented trajectories, and applying it to nine other trackers improved their IDF1 scores by 1 to 10 points. Accepted at ECCV 2022, the reference implementation builds on the YOLOX detector and reports 80.3 MOTA, 77.3 IDF1, and 63.1 HOTA on the MOT17 test set at 30 FPS on a single V100 GPU. The repository includes training on custom COCO-format datasets, evaluation tooling, and deployment paths for ONNX Runtime, TensorRT in Python and C++, ncnn, and NVIDIA DeepStream, and the method is re-implemented in Baidu's PaddleDetection. Released under the MIT license, it serves as a common baseline for researchers and engineers building pedestrian and vehicle tracking systems.