JPGNet: Joint Predictive Filtering and Generative Network for Image Inpainting

We propose a new method for image inpainting by formulating it as a mix of two problems that predictive filtering and deep generation. This work has been accepted to ACM-MM 2021. Please refer to the paper for details: https://arxiv.org/pdf/2107.04281.pdf

Dataset

Places2 Data of Places365-Standard
CelebA(https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html)
Dunhuang
Mask (https://nv-adlr.github.io/publication/partialconv-inpainting)

For data folder path (CelebA) organize them as following:

--CelebA
   --train
      --1-1.png
   --valid
      --1-1.png
   --test
      --1-1.png
   --mask-train
	  --1-1.png
   --mask-valid
      --1-1.png
   --mask-test
      --0%-20%
        --1-1.png
      --20%-40%
        --1-1.png
      --40%-60%
        --1-1.png

Run the code ./data/data_list.py to generate the data list

Pretrained models

We release our pretrained model (CelebA) at models

pretrained model (Places2) at models

pretrained model (Dunhuang) at models

Train

python train.py
For the parameters: checkpoints/config.yml, kpn/config.py

Test

python test.py
For the parameters: checkpoints/config.yml, kpn/config.py

Results

Comparsion with SOTA, see paper for details.

More details are coming soon

Bibtex

@article{guo2021jpgnet,
  title={JPGNet: Joint Predictive Filtering and Generative Network for Image Inpainting},
  author={Guo, Qing and Li, Xiaoguang and Juefei-Xu, Felix and Yu, Hongkai and Liu, Yang and others},
  journal={ACM-MM},
  year={2021}
}

Acknowledgments

Parts of this code were derived from:
https://github.com/tsingqguo/efficientderain
https://github.com/knazeri/edge-connect
https://github.com/RenYurui/StructureFlow
https://github.com/jingyuanli001/RFR-Inpainting

tsingqguo/jpgnet