This repository is the pytorch implementation of Progressive Image Inpainting with Full-Resolution Residual Network (arxiv)
We propose FRRN which shows that residual sturcture is particularly suitable for progressive inpainting strategy. Our spatial progressive inpainting model is composed of eight dilation modules. Each dilation module contains two full-resolution residual blocks. This architecture is designed to accurately control the mask-updating process and inpainting quality. Full-resolution branch is also helpful to determine the dilation stride of each FRRB and simultaneously improve final performance.
This is the dirty version for our implementation. Part of the code is modified from EdgeConnect. We note that partialconv2d.py is also modified based on the released version of PartialConv.
You can generate flist files of data through the script located in ./flist.
cd ./flist
bash flist.sh
You should change the data location in flist.sh first to generate your own flist files.
If you are willing to try different loss weights, you may get even higher PSNR value (e.g., you can decrease the weight of adversarial loss). However, we manage to achieve the balance between qualitative results and quantitative results. So we finally choose an appropriate setting of loss weights.
To train your own parameters of FRRN, you can enter
cd ./src
python main.py --gpu_id=0,1
To evaluate your trained model, you can enter
cd ./src
python main.py --skip_training --RESUME=True --gpu_id=0,1
We here provide our pre-trained model on Places2 dataset which performs closely to our reported statistics. We put it in ./save_models. If you use this pre-trained weights of model, you can
cd ./src
python main.py --skip_training --RESUME=True --gpu_id=0
Our code only resize input images to 256*256 resolution. So if you are curious about our network's performance on larger resolution, you should reset the related parameters and train your model from the beginning!
If you find these code is useful, don't hesitate to give star to this repository!!!! If you use it for your research, please cite our paper Progressive Image Inpainting with Full-Resolution Residual Network
@inproceedings{guo2019progressive,
title={Progressive Image Inpainting with Full-Resolution Residual Network},
author={Guo, Zongyu and Chen, Zhibo and Yu, Tao and Chen, Jiale and Liu, Sen},
booktitle={Proceedings of the 27th ACM International Conference on Multimedia},
pages={2496--2504},
year={2019},
organization={ACM}
}