170744039 / gsrnet

GSRNet

Code for the GSRNet (AAAI 2020)

Environment

tensorflow 1.4.0, python3.4, cuda 8.0.44 cudnn 6.0

Other packages please run:

pip install -r requirements.txt

Download ImageNet pre-trained model:

Refer to https://github.com/DrSleep/tensorflow-deeplab-lfov for more detail

Download model.ckpt-pretrained, net_skeleton.ckpt and put it in the 'ckpt' folder

Prepare tfrecords for dataset

1. Change the arguments of dataset, train_dir, mask_dir, output_directory, to corresponding directories.

2. Run python3 im_pre_casia_pair.py

3. The tfrecords for COCO and CASIA are provided in https://drive.google.com/drive/folders/1YY4UM1PBTbBWMyjx350ubp5udGZG66K1?usp=sharing

Train the model:

1. Change the tfrecords directory in train_default.sh
2. Run train_default.sh

Test the model

1. python3 dry_run.py --model_weights='./snapshots/$FIXME' --dataset='$FIXME'

for single image, use --dataset='single_img'

2. save output image: python3 dry_run.py --model_weights='./snapshots/$FIXME' --dataset='$FIXME' --save-dir='./output/$FIXME/' --vis=True

3. visualize generated images: python3 dry_run.py --model_weights='./snapshots/$FIXME' --dataset='$FIXME' --save-dir='./output/$FIXME/' --vis=True --vis_gan=True

Citation:

If this code or dataset helps your research, please cite our paper:

@inproceedings{zhou2020generate,
  title={Generate, Segment, and Refine: Towards Generic Manipulation Segmentation},
  author={Zhou, Peng and Chen, Bor-Chun and Han, Xintong and Najibi, Mahyar and Shrivastava, Abhinav and Lim, Ser Nam and Davis, Larry S},
  booktitle = {AAAI},
  year={2020}
}