attentive-gan-derainnet-pytorch

This is the pytorch implementation of CVPR2018 paper "Attentive Generative Adversarial Network for Raindrop Removal from A Single Image". You can refer to their paper for details [Paper Link]. This model consists of a attentive attentive-recurrent network, a contextual autoencoder network and a discriminative network. Using convolution lstm unit to generate attention map which is used to help locating the rain drop, multi-scale losses and a perceptual loss to train the context autoencoder network. Thanks for the origin author Rui Qian

If you find the resource useful, please cite the original paper and my repo attentive-gan-derainnet-pytorch

@InProceedings{Qian_2018_CVPR,
author = {Qian, Rui and Tan, Robby T. and Yang, Wenhan and Su, Jiajun and Liu, Jiaying},
title = {Attentive Generative Adversarial Network for Raindrop Removal From a Single Image},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}

Train Model

python train.py --seed randomseed --train_dir train_dir --val_dir validation_dir --batch_size bs --device device_num --checkpoint_every 500 --validate_every 500 --visualize_every 500

You may monitor the training process using tensorboard tools

During my experiment the loss G drops as follows:

The Image PSNR and SSIM between generated image and clean label image raises as follows:

The Image derain results as follows: (input, gt, att_map_0,..,att_map_4, generated)

Test Model

coming soon

TODO

• Prepare Inference Code