OC-SORT

This is the github repo for Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking [arxiv].

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.

News

[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

Dataset	HOTA	AssA	IDF1	MOTA	FP	FN	IDs	Frag
MOT17 (private)	63.2	63.2	77.5	78.0	15,129	107,055	1,950	2,040
MOT17 (public)	52.4	57.6	65.1	58.2	4,379	230,449	784	2,006
MOT20 (private)	62.4	62.5	76.4	75.9	20,218	103,791	938	1,004
MOT20 (public)	54.3	59.5	67.0	59.9	4,434	202,502	554	2,345
KITTI-cars	76.5	76.4	-	90.3	2,685	407	250	280
KITTI-pedestrian	54.7	59.1	-	65.1	6,422	1,443	204	609
DanceTrack	55.1	38.0	54.2	89.4	114,107	139,083	1,992	3,838
CroHD HeadTrack	44.1	-	62.9	67.9	102,050	164,090	4,243	10,122

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

Model Zoo

The pretrained model trained on DanceTrack-train set is the original version provided in DanceTrack, we rename and redirect it on Google Drive for convenience. The weights should output results as below:

Dataset	HOTA	IDF1	AssA	MOTA	DetA
DanceTrack-val	52.1	51.6	35.3	87.3	77.2
DanceTrack-test	55.1	54.2	38.0	89.4	80.3

For model weights for other datasets, please refer to the model zoo of ByteTrack for options.

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/bytetrack_dance_model.pth.tar --path videos/dance_demo.mp4 --fp16 --fuse --save_result --out_path demo_out.mp4

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:

@article{cao2022observation,
  title={Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking},
  author={Cao, Jinkun and Weng, Xinshuo and Khirodkar, Rawal and Pang, Jiangmiao and Kitani, Kris},
  journal={arXiv preprint arXiv:2203.14360},
  year={2022}
}