Deep-OC-SORT

Deep OC-SORT: Multi-Pedestrian Tracking by Adaptive Re-Identification [arxiv]

Gerard Maggiolino*, Adnan Ahmad*, Jinkun Cao, Kris Kitani (*=equal contribution)

Dataset	HOTA	AssA	IDF1	MOTA	IDs	Frag
MOT17	64.9	65.9	80.6	79.4	1,950	2,040
MOT20	63.9	65.9	79.2	75.6	779	1,536

Dataset	HOTA	AssA	DetA	MOTA	IDF1
DanceTrack	61.3	45.8	82.2	92.3	61.5

• As of Mar 9th, 2023, Deep-OC-SORT ranks 1st compared to published methods on MOT17 and MOT20 w.r.t. HOTA. It improves tracking performance on DanceTrack over OC-SORT by ~6 HOTA.

Installation

Tested with Python3.8 on Ubuntu 18.04. More versions will likely work.

After cloning, install external dependencies:

cd external/YOLOX/
pip install -r requirements.txt && python setup.py develop
cd ../external/deep-person-reid/
pip install -r requirements.txt && python setup.py develop
cd ../external/fast_reid/
pip install -r docs/requirements.txt

OCSORT dependencies are included in the external dependencies. If you're unable to install faiss-gpu needed by fast_reid, faiss-cpu should be adequate. Check the external READMEs for any installation issues.

Add the weights to the external/weights directory (do NOT untar the .pth.tar YOLOX files).

Data

Place MOT17/20 and DanceTrack under:

data
|——————mot (this is MOT17)
|        └——————train
|        └——————test
|——————MOT20
|        └——————train
|        └——————test
|——————dancetrack
|        └——————train
|        └——————test
|        └——————val

and run:

python3 data/tools/convert_mot17_to_coco.py
python3 data/tools/convert_mot20_to_coco.py
python3 data/tools/convert_dance_to_coco.py

Evaluation

For the MOT17/20 and DanceTrack baseline:

exp=baseline
# Flags to disable all the new changes
python3 main.py --exp_name $exp --post --emb_off --cmc_off --aw_off --new_kf_off --grid_off --dataset mot17
python3 main.py --exp_name $exp --post --emb_off --cmc_off --aw_off --new_kf_off --grid_off -dataset mot20 --track_thresh 0.4
python3 main.py --exp_name $exp --post --emb_off --cmc_off --aw_off --new_kf_off --grid_off --dataset dance --aspect_ratio_thresh 1000

This will cache detections under ./cache, speeding up future runs. This will create results at:

# For the standard results
results/trackers/<DATASET NAME>-val/$exp.
# For the results with post-processing linear interpolation
results/trackers/<DATASET NAME>-val/${exp}_post.

To run TrackEval for HOTA and Identity with linear post-processing on MOT17, run:

python3 external/TrackEval/scripts/run_mot_challenge.py \
  --SPLIT_TO_EVAL val \
  --METRICS HOTA Identity \
  --TRACKERS_TO_EVAL ${exp}_post \
  --GT_FOLDER results/gt/ \
  --TRACKERS_FOLDER results/trackers/ \
  --BENCHMARK MOT17

Replace that last argument with MOT17 / MOT20 / DANCE to evaluate those datasets.

For the highest reported ablation results, run:

exp=best_paper_ablations
python3 main.py --exp_name $exp --post --grid_off --new_kf_off --dataset mot17 --w_assoc_emb 0.75 --aw_param 0.5
python3 main.py --exp_name $exp --post --grid_off --new_kf_off --dataset mot20 --track_thresh 0.4 --w_assoc_emb 0.75 --aw_param 0.5
python3 main.py --exp_name $exp --post --grid_off --new_kf_off --dataset dance --aspect_ratio_thresh 1000 --w_assoc_emb 1.25 --aw_param 1

This will cache generated embeddings under ./cache/embeddings, speeding up future runs. Re-run the TrackEval script provided above.

You can achieve higher results on individual datasets with different parameters, but we kept them fairly consistent with round numbers to avoid over-tuning.

Contributing

Formatted with black --line-length=120 --exclude external .

Citation

If you find our work useful, please cite our paper:

@article{maggiolino2023deep,
    title={Deep OC-SORT: Multi-Pedestrian Tracking by Adaptive Re-Identification}, 
    author={Maggiolino, Gerard and Ahmad, Adnan and Cao, Jinkun and Kitani, Kris},
    journal={arXiv preprint arXiv:2302.11813},
    year={2023},
}

Also see OC-SORT, which we base our work upon:

@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}
}