OC-SORT

Observation-Centric SORT (OC-SORT) is a pure motion-model-based multi-object tracker. It aims to improve tracking robustness in crowded scenes and when objects are in non-linear motion. It is designed by recognizing and fixing limitations in Kalman filter and SORT. It is flexible to integrate with different detectors and matching modules, such as appearance similarity. It remains, Simple, Online and Real-time.

Pipeline

Observation-centric Re-Update

News

• [07/09/2023]: A C++ support is provided. See the doc for instructions. Thanks for the contribution!
• [07/01/2023]: Deep OC-SORT is accepted to ICIP2023. It adds an adaptive appeareance similarity-based association upon OC-SORT.
• [03/15/2023]: We update the preprint version on Arxiv. We rename OOS to be "Observation-centric Re-Update" (ORU).
• [02/28/2023]: OC-SORT is accepted to CVPR 2023. We will update the code and paper soon. We made intensive revision of the paper writing.
• [02/26/2023]: Deep-OC-SORT, a combination of OC-SORT and deep visual appearance, is released on Github and Arxiv. Significant performance improvement on MOT17, MOT20 and DanceTrack.
• [08/16/2022]: Support OC-SORT in mmtracking. If you want to do tracking with more advanced and customizable experience, you may want to give it a try. The mmtracking version is still in-preview. Performance on more datasets to be verified.
• [04/27/2022]: Support intergration with BYTE and multiple cost metrics, such as GIoU, CIoU, etc.
• [04/02/2022]: A preview version is released after a primary cleanup and refactor.
• [03/27/2022]: The arxiv preprint of OC-SORT is released.

Benchmark Performance

• Results are from reusing detections of previous methods and shared hyper-parameters. Tune the implementation adaptive to datasets may get higher performance.
• The inference speed is ~28FPS by a RTX 2080Ti GPU. If the detections are provided, the inference speed of OC-SORT association is 700FPS by a i9-3.0GHz CPU.

• A sample from DanceTrack-test set is as below and more visualizatiosn are available on Google Drive

  

Get Started

• See INSTALL.md for instructions of installing required components.

• See GET_STARTED.md for how to get started with OC-SORT.

• See MODEL_ZOO.md for available YOLOX weights.

• See DEPLOY.md for deployment support over ONNX, TensorRT and ncnn.

Demo

To run the tracker on a provided demo video from Youtube:

python3 tools/demo_track.py --demo_type video -f exps/example/mot/yolox_dancetrack_test.py -c pretrained/ocsort_dance_model.pth.tar --path videos/dance_demo.mp4 --fp16 --fuse --save_result --out_path demo_out.mp4

Roadmap

We are still actively updating OC-SORT. We always welcome contributions to make it better for the community. We have some high-priorty to-dos as below:

• [x] Add more asssocitaion cost choices: GIoU, CIoU, etc.
• [x] Support OC-SORT in mmtracking.
• [ ] Add more deployment options and improve the inference speed.
• [x] Make OC-SORT adaptive to customized detector (in the mmtracking version).

Acknowledgement and Citation

The codebase is built highly upon YOLOX, filterpy, and ByteTrack. We thank their wondeful works. OC-SORT, filterpy and ByteTrack are available under MIT License. And YOLOX uses Apache License 2.0 License.

If you find this work useful, please consider to cite our paper:

@inproceedings{cao2023observation,
  title={Observation-centric sort: Rethinking sort for robust multi-object tracking},
  author={Cao, Jinkun and Pang, Jiangmiao and Weng, Xinshuo and Khirodkar, Rawal and Kitani, Kris},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9686--9696},
  year={2023}
}