Overview
ByteTrack is a multi-object tracking (MOT) method that associates every detection box, including those with low scores, to recover true objects and filter out background detections. This approach addresses the problem of missing objects and fragmented trajectories caused by discarding low-score detection boxes in traditional MOT methods. ByteTrack demonstrates significant improvements in IDF1 scores when applied to various state-of-the-art trackers and achieves high MOTA, IDF1, and HOTA scores on the MOT17 test set. Implemented using YOLOX for detection, it provides demo links for Google Colab and Huggingface Spaces. The tracker can be installed and used on a host machine or via Docker, with detailed instructions provided for data preparation and model training, supporting datasets like MOT17, MOT20, CrowdHuman, Cityperson, and ETHZ.