Authors : Mehdi Miah, Justine Pepin, Nicolas Saunier & Guillaume-Alexandre Bilodeau
Polytechnique Montréal - 2020

[Project page] [arXiv paper] [ICPR 2020 paper]

Rank visual features for multiple object tracking focused on urban scenes

Linux and Windows supported. Python 3.6, Pytorch 1.4, CUDA 10.0

cd src
git clone https://github.com/KaiyangZhou/deep-person-reid.git #(torchreid 1.2.7)
mv deep-person-reid/ deep_person_reid/
pip install efficientnet_pytorch
git clone https://github.com/cw1204772/AIC2018_iamai.git

You can download the following datasets :

• WildTrack : https://www.epfl.ch/labs/cvlab/data/data-wildtrack/
• MOT17 : https://motchallenge.net/data/MOT17/
• DETRAC : http://detrac-db.rit.albany.edu/
• UAVDT : https://sites.google.com/site/daviddo0323/projects/uavdt

Change the path to data in the file ./src/dataset.py.

The weights for VGG-19, ResNet-18, DenseNet-121 come from Pytorch.

The weights for Efficient-B0 come from efficientnet_pytorch (automatically downloaded).

The weights for pedestrian ReID come from deep-person-reid

The weights for vehicles ReID come from AIC2018_iamai

.
├─── doc
├─── results
│   ├─── figure
│   ├─── logs
├─── script
│   ├─── run_DETRAC.sh
│   ├─── run_MOT17.sh
│   ├─── run_UAVDT.sh
│   ├─── run_WildTrack.sh
├─── src
│   ├─── affinity.py
│   ├─── analysis_rank.py
│   ├─── analysis_size.py
│   ├─── appearances.py
│   ├─── dataset.py
│   ├─── main.py
│   ├─── utils.py
│   ├─── AIC2018_iamai
│   ├─── deep_person_reid
├─── weights
│   ├─── model_880_base.ckpt
│   ├─── osnet_ain_x1_0_market1501_256x128_amsgrad_ep100_lr0.0015_coslr_b64_fb10_softmax_labsmth_flip_jitter.pth

Compute the mean average precision on a specific scene with a chosen features and sigma :

python main.py --dataset=DETRAC --scene=20011 --feature=resnet18 --sigma=10

Compute the mean average precision on a whole dataset :

cd src ; bash ../script/run_DETRAC.sh

Analyze the ranking of features on a specific dataset :

python analysis_rank.py --dataset=DETRAC

Analyze effect of the size on features on a specific dataset :

python analysis_size.py --dataset=DETRAC

If you refer to this work, please cite :

@inproceedings{miah2020empirical,
    title = {An {Empirical} {Analysis} of {Visual} {Features} for {Multiple} {Object} {Tracking} in {Urban} {Scenes}},
    author = {Miah, Mehdi and Pepin, Justine and Saunier, Nicolas and Bilodeau, Guillaume-Alexandre},
    booktitle = {International {Conference} on {Pattern} {Recognition} ({ICPR})},
    year = {2020}
}

We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [CRDPJ 528786 - 18], [DG 2017-06115] and the support of Arcturus Networks.