Home

Awesome

CVPR2020: How To Train Your Deep Multi-Object Tracker <br />

License: LGPL v3

News: We release the code for training and testing DeepMOT-Tracktor and the code for training DHN. Please visit: https://gitlab.inria.fr/robotlearn/deepmot

How To Train Your Deep Multi-Object Tracker <br /> Yihong Xu, Aljosa Osep, Yutong Ban, Radu Horaud, Laura Leal-Taixé, Xavier Alameda-Pineda <br /> [Paper] <br />

Bibtex

If you find this code useful, please star the project and consider citing:

@inproceedings{xu2020train,
  title={How To Train Your Deep Multi-Object Tracker},
  author={Xu, Yihong and Osep, Aljosa and Ban, Yutong and Horaud, Radu and Leal-Taix{\'e}, Laura and Alameda-Pineda, Xavier},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6787--6796},
  year={2020}
}
<div align="center"> <img src="https://gitlab.inria.fr/yixu/deepmot/-/raw/master/teaser.pdf" width="800px" /> </div>

Environment setup <a name="environment-setup">

This code has been tested on Ubuntu 16.04, Python 3.6, Pytorch=0.4.1, CUDA 9.2, GTX 1080Ti, Titan X, and RTX Titan GPUs.

Warning: the results can be slightly different due to Pytorch version and CUDA version.

git clone git@gitlab.inria.fr:yixu/deepmot.git && cd deepmot

Option 1:

Option 2 (recommended):

we provide a Singularity image with all packages pre-installed (similar to Docker) for training and testing.

singularity shell --nv --bind yourLocalPath:yourPathInsideImage tracker.sif

- -bind: to link a singularity path with a local path. By doing this, you can find data from local PC inside Singularity image; <br /> - -nv: use the local Nvidia driver.

Testing <a name="testing-models">

dataset: mot17_train_17

For MOT16 (images as the same as MOT17):

dataset: mot17_all_DPM_RAW16
python test_tracktor/experiments/scripts/tst_tracktor_private.pytst_tracktor_pub_reid.py (public detections) or test_tracktor/experiments/scripts/tst_tracktor_private.py (private detections)

The results are saved by default under test_tracktor/output/log/, you can modify it by changing output_dir in the test_tracktor/experiments/cfgs/tracktor_pub_reid.yaml and test_tracktor/experiments/cfgs/tracktor_private.yaml.

Training <a name="training-models">

python train_tracktor/experiments/scripts/train_tracktor_full.py

The trained models are saved by default under train_tracktor/output/log_full/ folder. <br /> The tensorboard logs are saved by default under deepmot/logs/train_log/ folder and you can visualize your training process by:

tensorboard --logdir=YourGitFolder/train_tracktor/output/log_full/

Note:

pip install --upgrade tensorflow

Train DHN

python train_DHN/train_DHN.py --is_cuda --bidirectional

for more parameter details please run:

python train_DHN/train_DHN.py -h

By default the trained models are saved into train_DHN/output/DHN/ and log files in train_DHN/log/

your can visualize the training via tensorboard:

tensorboard --logdir=YourGitFolder/train_DHN/log/

Note:

pip install --upgrade tensorflow

Evaluation

You can run test_tracktor/experiments/scripts/evaluate.py to evaluate your tracker's performance.

python test_tracktor/experiments/scripts/evaluate.py

Results

MOT17 public detections:

datasetMOTAMOTPFNFPIDsWTotal Nb. Objs
train62.5%91.7%124786887798336891
test53.7%77.2%247447117311947564228

MOT16 public detections:

datasetMOTAMOTPFNFPIDsWTotal Nb. Objs
train58.8%92.2%44711538229110407
test54.8%77.5%787652955645182326

MOT16/17 private detections:

datasetMOTAMOTPFNFPIDsWTotal Nb. Objs
train70.0%91.3%32513552677112297

Note:

Demo <a name="demo">

<div align="center"> <img src="https://github.com/yihongXU/deepMOT/raw/master/trackingresults.gif" width="500px" /> </div>

Acknowledgement <a name="Acknowledgement">

Some code is modified and network pre-trained weights are obtained from the following repositories: <br />

Single Object Tracker: SiamRPN, Tracktor, Faster-RCNN pytorch implementation.

@inproceedings{Zhu_2018_ECCV,
  title={Distractor-aware Siamese Networks for Visual Object Tracking},
  author={Zhu, Zheng and Wang, Qiang and Bo, Li and Wu, Wei and Yan, Junjie and Hu, Weiming},
  booktitle={European Conference on Computer Vision},
  year={2018}
}

@InProceedings{Li_2018_CVPR,
  title = {High Performance Visual Tracking With Siamese Region Proposal Network},
  author = {Li, Bo and Yan, Junjie and Wu, Wei and Zhu, Zheng and Hu, Xiaolin},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2018}
}

@InProceedings{tracktor_2019_ICCV,
  author = {Bergmann, Philipp and Meinhardt, Tim and Leal{-}Taix{\'{e}}}, Laura},
  title = {Tracking Without Bells and Whistles},
  booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
  month = {October},
  year = {2019}}

@inproceedings{10.5555/2969239.2969250,
author = {Ren, Shaoqing and He, Kaiming and Girshick, Ross and Sun, Jian},
title = {Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks},
year = {2015},
publisher = {MIT Press},
address = {Cambridge, MA, USA},
booktitle = {Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1},
pages = {91–99},
numpages = {9},
location = {Montreal, Canada},
series = {NIPS’15}
}

MOT Metrics in Python: py-motmetrics<br /> Appearance Features Extractor: DAN<br />

@article{sun2018deep,
  title={Deep Affinity Network for Multiple Object Tracking},
  author={Sun, ShiJie and Akhtar, Naveed and Song, HuanSheng and Mian, Ajmal and Shah, Mubarak},
  journal={arXiv preprint arXiv:1810.11780},
  year={2018}
}

Training and testing Data from: <br /> MOT Challenge: motchallenge

@article{MOT16,
    title = {{MOT}16: {A} Benchmark for Multi-Object Tracking},
    shorttitle = {MOT16},
    url = {http://arxiv.org/abs/1603.00831},
    journal = {arXiv:1603.00831 [cs]},
    author = {Milan, A. and Leal-Taix\'{e}, L. and Reid, I. and Roth, S. and Schindler, K.},
    month = mar,
    year = {2016},
    note = {arXiv: 1603.00831},
    keywords = {Computer Science - Computer Vision and Pattern Recognition}
}