CenterTrack

CenterTrack: Tracking Objects as Points

Citation

The code is built with CenterTrack. Thanks for their great work.

Installation

Please refer to INSTALL.md for installation instructions.

Data preparation

The data in the following structure:

DIVOTrack
    └——————datasets
    |        └——————DIVO
    |           |——————images
    |           |        └——————train
    |           |        └——————test
    |           └——————labels_with_ids
    |           |        └——————train
    |           |        └——————test  
    |           └——————annoatations
    |           |        └——————train.json
    |           |        └——————test.json
    └——————${ROOT}

If your data does not has the "annotations" directory, please use the following command to generate

cd ${CenterTrack_ROOT}
cd src
python convert_divo_to_coco.py

Train

The pre-trained model can download from Google Drive and put it into ./models/

cd ${CenterTrack_ROOT}
cd src
python main.py tracking --exp_id divo --dataset divo --pre_hm --ltrb_amodal --same_aug --hm_disturb 0.05 --lost_disturb 0.4 --fp_disturb 0.1 --gpus 0,1 --load_model ../models/crowdhuman.pth --num_epochs 30

or

sh ./experiments/divo_train.sh

Inference

Make sure the "exp_id" is the same as the training one, the model will be directly loaded from the corresponding dir "exp/tracking/"exp_id"/model_last.pth".

cd ${CenterTrack_ROOT}
cd src
python test.py tracking --exp_id divo --dataset divo --pre_hm --ltrb_amodal --track_thresh 0.4 --pre_thresh 0.5 --resume

or

sh ./experiments/divo_test.sh

The result will be saved to "exp/tracking/"exp_id"/result_divo

Evaluation

0. Download our final model and put it into ./exp/tracking/"exp_id"/
1. Change the directory name from "result_divo" to "centertrack"
2. Make sure "centertrack" has the middle directory "data". (i.e. centertrack/data/circleRegion_Drone.txt instead of centertrack/circleRegion_Drone.txt)
3. Copy your result_divo to DIVOTrack/TrackEval/data/trackers/mot_challenge/divo
4. Go to DIVOTrack/TrackEval
5. See the instrcution on TrackEval

License

CenterTrack is developed upon CenterNet. Both codebases are released under MIT License themselves. Some code of CenterNet are from third-parties with different licenses, please check the CenterNet repo for details. In addition, this repo uses py-motmetrics for MOT evaluation and nuscenes-devkit for nuScenes evaluation and preprocessing. See NOTICE for detail. Please note the licenses of each dataset. Most of the datasets we used in this project are under non-commercial licenses.