/ST-SACLF

Primary LanguagePython

Painting Classification with Style Transfer

Code for the VISAPP'23 paper 'Tackling Data Bias in Painting Classification with Style Transfer'.

Style Transfer-Spatial Attention CLassiFier

Model performance

Spatial Attention maps

Dataset

The datasets can be downloaded using the link:

https://drive.google.com/drive/folders/1j7nIQzmoXBm2GLPt3EJ0odl8IvLZgizo?usp=share_link

unzip kaokore_control_v1.zip

unzip kaokore_imagenet_style.zip

Requirements

Pytorch

Pytorch Lightning

Wandb

Usage

The datasets should be outside the repository if downloaded from the google drive link.

Style transfer to generate the augmented dataset

python pytorch-AdaIN/train.py --save_dir models --max_iter 20000

python pytorch-AdaIN/test.py --decoder models/decoder_iter_20000_kaokore_stylized.pth

Training and testing the model

python train_test.py

Credits

attention.py has been adapted from https://github.com/0aqz0/pytorch-attention-mechanism.git

pytorch-AdaIN has been adapted from https://github.com/naoto0804/pytorch-AdaIN.git

Citation

If you use our code or data, please cite:

@inproceedings{mridula2023tackling,
    title={Tackling Data Bias in Painting Classification with Style Transfer},
    author={Vijendran, Mridula and Li, Frederick W. B. and Shum, Hubert P. H.},
    booktitle={International Conference on Computer Vision Theory and Applications (VISAPP)},
    year={2023}
}