Image Harmonization with Attention-based Deep Feature Modulation

This code provides a pytorch implementation of "Image Harmonization with Attention-based Deep Feature Modulation".

Requirements

Python 3.7
PyTorch tested on 1.4.0
PIL
skimage
tqdm
numpy

Dataset

Download the iHarmony4 dataset from this link.

We resized all training images of HAdobe5k subset with a max size of 1024. Original images of HAdobe5k are significantly large, which slow down the data loading process. In order to speed up this process, we resize all training images of HAdobe5k dataset.

To resize the training images of HAdobe5k, specify the paths in the following script and run the script.

python resize.py

Training

Specify paths in the train.sh script before runing the script.

bash train.sh

Testing

Here we provide a better model compared to the model used in our paper. Although the performance on the 'ALL' dataset is the same as in the paper, this model outperforms previous methods on all sub-datasets.

To download the model, run the following script.

bash download_model.sh

To test the performances, run the following script.

bash test.sh

Citation

If you find the code useful in your research, please consider citing our paper:

@InProceedings{Hao2020bmcv,
author       = "Guoqing Hao and Satoshi Iizuka and Kazuhiro Fukui",
title        = "Image Harmonization with Attention-based Deep Feature Modulation",
booktitle    = "The British Machine Vision Conference (BMCV)",
year         = "2020",
}