LBAM_inpainting

Introduction

This is the pytorch implementation of Paper: Image Inpainting With Learnable Bidirectional Attention Maps (ICCV 2019) paper suppl

Model Architecture

We propose a Bidirectional Attention model based on the U-Net architecture.

Bidrectional Attention Layer

Prerequisites

• Python 3.6
• Pytorch >= 1.0 (tested on pytorch version 1.0.0, 1.2.0, 1.3.0)
• CPU or NVIDIA GPU + Cuda + Cudnn

Training

To train the LBAM model:

python train.py --batchSize numOf_batch_size --dataRoot your_image_path \
--maskRoot your_mask_root --modelsSavePath path_to_save_your_model \
--logPath path_to_save_tensorboard_log --pretrain(optional) pretrained_model_path

Testing

To test the model:

python test.py --input input_image --mask your_mask --output output_file_prefix --pretrain pretrained_model_path

To test with random batch with random masks:

python test_random_batch.py --dataRoot your_image_path
--maskRoot your_mask_path --batchSize numOf_batch_size --pretrain pretrained_model_path

Some Results

We suggest that you train our model with a large batch size (>= 48 or so). We re-train our model with batch size 10, the results degrades a little bit, I guess it may be due to the batch-normalization opreation (I would try removing bn from LBAM and see how it affects).

The pretrained model can be found at google drive, or baidu cloud with extract code: mvzh. I made a slight change by setting the bn to false and modify the last tanh from absolute value to (tanh() + 1) / 2.

Here are some inpainting results that we train with batch size of 10 on Paris StreetView dataset:

Input	Results	Ground-Truth

If you find this code would be useful

Please cite our paper

@InProceedings{Xie_2019_ICCV,
author = {Xie, Chaohao and Liu, Shaohui and Li, Chao and Cheng, Ming-Ming and Zuo, Wangmeng and Liu, Xiao and Wen, Shilei and Ding, Errui},
title = {Image Inpainting With Learnable Bidirectional Attention Maps},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}

Acknowledgement

We benifit a lot from NVIDIA-partialconv and naoto0804-pytorch-inpainting-with-partial-conv, thanks for their excellent work.