luckynote / deepMOT

Official implementation of DeepMOT: A Differentiable Framework for Training Multiple Object Trackers.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

DeepMOT

License: LGPL v3 HitCount

This is the official implementation with training code for the paper:

DeepMOT: A Differentiable Framework for Training Multiple Object Trackers
Yihong Xu, Yutong Ban, Xavier Alameda-Pineda, Radu Horaud
[Paper]

Bibtex

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

@article{xu2019deepmot,
  title={DeepMOT: A Differentiable Framework for Training Multiple Object Trackers},
  author={Xu, Yihong and Ban, Yutong and Alameda-Pineda, Xavier and Horaud, Radu},
  journal={arXiv preprint arXiv:1906.06618},
  year={2019}
}

Contents

  1. Environment Setup
  2. Testing
  3. Training
  4. Demo
  5. Acknowledgement

Environment setup

This code has been tested on Ubuntu 16.04, Python 3.6, Pytorch> 0.4.1, CUDA 9.0 and CUDA 10.0, GTX 1080Ti, Titan X and RTX Titan GPUs.

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

  • Clone the repository
git clone https://github.com/yihongXU/deepMOT.git && cd deepmot

Option 1:

  • Setup python environment
conda create -n deepmot python=3.6
source activate deepmot
pip install -r requirements.txt

Option 2: we offer Singularity images (similar to Docker) for training and testing.

  • Open a terminal
  • Install singularity
sudo apt-get install -y singularity-container
cd deepmot
singularity shell --nv --bind yourLocalPath:yourPathInsideImage ./SingularityImages/pytorch1-1-cuda100-cudnn75.simg

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

Testing

We provide code for performing tracking with our pre-trained models on MOT Challenge dataset. The code outputs txt files for MOT Challenge submissions, they can also be used for plotting bounding boxes and visualization.

  • Setup your environment

  • Download MOT data Dataset can be downloaded here: e.g. MOT17

  • Put MOT dataset into deepmot/data/ and it should have the following structure:

            mot
            |-------train
            |    |
            |    |---video_folder1
            |    |   |---det
            |    |   |---gt
            |    |   |---img1
            |    |
            |    |---video_folder2
            ...
            |-------test
            |    |
            |    |---video_folder1
            |    |   |---det
            |    |   |---img1
            ...

-Put all pre-trained models to deepmot/pretrained/

  • run tracking code
python tracking_on_mot.py

for more details about parameters, do:

python tracking_on_mot.py -h

The results are save by default under deepmot/saved_results/txts/test_folder/.

  • Visualization After finishing tracking, you can visualize your results by plotting bounding box to images.
python plot_results.py

the results are save by default under deepmot/saved_results/imgs/test_folder

Note:

  • we clean the detections with nms and threshold of detection scores. They are saved into numpy array in the folder deepmot/clean_detections, if you have trouble opening them, try to add allow_pickle=True to np.load() function.

Results

We provide codes for evaluting tracking results in terms of MOTP and MOTA:

python evaluation.py --txts_path=yourTxTfilesFolder

MOT17:

dataset MOTA MOTP FN FP IDsW Total Nb. Objs
train 49.249% 82.812% 149575 19807 1592 336891
test 48.500% 76.900% 262765 24544 3160 564228

Note:

  • the results are better than reported in the paper because we add Camera Motion Compensation to deal with moving camera videos.
  • the results can be slightly different depending on the running environment.

Training

  • Setup your environment

  • Download MOT data Dataset can be downloaded here: e.g. MOT17

  • Put MOT dataset into deepmot/data and it should have the following structure:

            mot
            |-------train
            |    |
            |    |---video_folder1
            |    |   |---det
            |    |   |---gt
            |    |   |---img1
            |    |
            |    |---video_folder2
            ...
            |-------test
            |    |
            |    |---video_folder1
            |    |   |---det
            |    |   |---img1
            ...

-Put SiamRPNVOT.model to deepmot/pretrained/ folder

  • run training code
python train_mot.py

for more details about parameters, do:

python train_mot.py -h

The trained models are save by default under deepmot/saved_models/ folder.
The tensorboard logs are saved by default under deepmot/logs/train_log/ folder and you can visualize your training process by:

tensorboard --logdir=/mnt/beegfs/perception/yixu/opensource/deepMOT/logs/train_log

Note:

pip install --upgrade tensorflow

Demo

Acknowledgement

Some codes are modified and network pretrained weights are obtained from the following repositories:
Single Object Tracker: SiamRPN

@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}
}

MOT Metrics in Python: py-motmetrics
Appearance Features Extractor: DAN

@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:
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}
}

About

Official implementation of DeepMOT: A Differentiable Framework for Training Multiple Object Trackers.

License:GNU Lesser General Public License v3.0


Languages

Language:Python 100.0%