Semantics-guided Part Attention Network

This is the pytorch implementatin of Semantics-guided Part Attention Network (SPAN)

Paper

Orientation-aware Vehicle Re-identification with Semantics-guided Part Attention Network
Tsai-Shien Chen, Chih-Ting Liu, Chih-Wei Wu, and Shao-Yi Chien
European Conference on Computer Vision (ECCV), Oral, 2020
[Paper Website] [arXiv]

Citation

If you use SPAN, please cite this paper:

@inproceedings{SPAN,
    title        = {Orientation-aware Vehicle Re-identification with Semantics-guided Part Attention Network},
    author       = {Chen, Tsai-Shien and Liu, Chih-Ting and Wu, Chih-Wei and Chien, Shao-Yi},
    booktitle    = {European Conference on Computer Vision},
    pages        = {330--346},
    year         = {2020},
    organization = {Springer}
}

Visualization Example

We visiualize some examples of vehicle images and their

• foreground masks generated by grabcut,
• foreground masks generated by deep learning network,
• front, rear and side attention masks.

Get Started

Prerequisites

• Download VeRi-776 dataset [link]
  Please note that our code is only supported on VeRi-776 dataset currently.
• Download this repo by:

$ git clone https://github.com/tsaishien-chen/SPAN.git
$ cd SPAN

• We run the code under Python 3.6.9
• The versions of used python packages are listed in requirements.txt. You can install all the dependancies by:

$ sed -i 's/==/>=/g' requirements.txt
$ pip3 install -r requirements.txt

Train

To train the model from scratch, please run

$ python3 main.py --mode train --image_root <Path_to_VeRi>

For example,

$ python3 main.py --mode train --image_root ../Dataset/VeRi

The whole training process includes five steps:

1. generating the foreground masks by grabcut,
2. training network to generate more robust foreground mask,
3. generating the foreground masks by deep learning network,
4. training network to generate part (front, rear and side) attention mask, and
5. generating part attention masks.

Implement

We have given the pretrained model of part attention generator;
therefore, you can simply generate the part attention mask without training by

$ python3 main.py --mode implement --image_root <Path_to_VeRi>

For example,

$ python3 main.py --mode implement --image_root ../Dataset/VeRi

Visualize

After training and implementation process, the code will automatically visualize generated masks as above.
Or, you can uncomment the visualize function in main.py and can independently visualize the masks after being generated in each step.

Contact

Tsai-Shien Chen, Media IC and System Lab, National Taiwan University
E-mail : tschen@media.ee.ntu.edu.tw