liuqk3/OUTrack

Introduction

This is the official implementation of Online Multi-Object Tracking with Unsupervised Re-Identification Learning and Occlusion Estimation. Note that:

• Because I am quite busy recently, the codes are prepared in a hurry, so they may look ugly.
• The work has been done in Seqtember 2020, so some hyper-parameters maybe not the same with those in the paper.

I will try to make the codes more accurate and beautiful when I am available.

Installation

• Clone this repo, and we'll call the directory that you cloned as ${OUTrack}
• Install dependencies. We use python 3.7 and pytorch >= 1.2.0

conda create -n OUTrack python=3.7
conda activate OUTrack
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch
cd ${OUTrack}
pip install -r requirements.txt

For windows users, there maybe some errors. What you need to do is to comment cython-box in requirements.txt. Then do as follows:

pip install -r requirements.txt
pip install numpy==1.19.3
somewhere=path/to/somewhere
cd $somewhere
git clone https://github.com/samson-wang/cython_bbox.git
cd cython_bbox
python setup.py build install

• We use DCNv2 in our backbone network and more details can be found in their repo.

# git clone https://github.com/CharlesShang/DCNv2
cd DCNv2
./make.sh

• In order to run the code for demos, you also need to install ffmpeg.

Data preparation

The CrowdHuman dataset is used for pretraining, it can be downloaded from their official webpage. After downloading, you should prepare the data in the following structure:

${OUTrack}/data/crowdhuman
   |——————images
   |        └——————train
   |        └——————val
   └------annotation_train.odgt
   └------annotation_val.odgt

We use MOTChallenge for training and evaluation. MOT16, MOT17 and MOT20 can be downloaded from the official webpage of MOT challenge. After downloading, you should prepare the data in the following structure:

${OUTrack}/data/MOT16
   |——————images
            └——————train
            └——————test

${OUTrack}/data/MOT17
   |——————images
            └——————train
            └——————test
${OUTrack}/data/MOT20
   |——————images
            └——————train
            └——————test

Then, you can run following command to get json style annotations:

cd src/tools
python gen_json_annotations.py

Pretrained models and baseline model

Coming soon ...

Training and Inference

• Download DLA-34 COCO pretranined CenterNet model weight from here, and put it to OUTrack/models/ctdet_coco_dla_2x.pth

• See mot17.sh for example. Note the commands in different bash scripts are not exactly the same with those used in the paper.

Note that the --debug option is quite useful for debugging while training or visaulize the tracking results online! If you are in debug (i.e. opt.debug > 0) mode, you can press Esc to stop.

Acknowledgement

A large part of the code is borrowed from FairMOT and CenterTrack. Thanks for their wonderful works.

Citation

@article{liu2022online,
  title={Online Multi-Object Tracking with Unsupervised Re-Identification Learning and Occlusion Estimation},
  author={Liu, Qiankun and Chen, Dongdong and Chu, Qi and Yuan, Lu and Liu, Bin and Zhang, Lei and Yu, Nenghai},
  journal={Neurocomputing},
  year={2022},
  publisher={Elsevier}
}