Towards Realtime MOT is the official open-source code for JDE — Joint Detection and Embedding — a multi-object tracker published on GitHub that detects people and tracks them in a single network, in near real time. It has around 2,400 GitHub stars.
Most trackers run a detector and a separate re-identification model; JDE learns both tasks together in one network, outputting bounding boxes and appearance embeddings in a single forward pass. That design made it one of the fastest accurate trackers of its era — the fastest released model runs at around 38 FPS — and the paper was accepted to ECCV 2020. JDE’s joint-learning idea influenced a generation of later trackers that adopted the same one-shot design. The repo includes training and evaluation code, a dataset zoo covering standard MOT benchmarks, pretrained models and a demo. The codebase was last updated in 2021 and has since been re-implemented by the community, including a pure C++ version and a PaddleDetection port, but this remains the canonical source.
Towards Realtime MOT is free and open source under the MIT license. Pretrained weights and training data links are provided in the repository at no cost; you supply your own GPU for training and inference.
Computer-vision engineers and researchers working on pedestrian tracking, surveillance analytics and sports or traffic video analysis, and students studying tracking-by-detection.