Wide-Context Semantic Image Extrapolation

by Yi Wang, Xin Tao, Xiaoyong Shen, Jiaya Jia.

Introduction

This repository gives the Tensorflow implementation of the method in CVPR 2019 paper, 'Wide-Context Semantic Image Extrapolation'. This method can expand semantically sensitive objects (face, body) / scenes beyond image boundary.

Partial Results

Faces

Bodies

Scenes

More results

Network structure

Key components

• Small-to-large scheme
• Context normalization
• Relative spatial variant loss

Prerequisites

• Python3.5 (or higher)
• Tensorflow 1.6 (or later versions, excluding 2.x) with NVIDIA GPU or CPU
• OpenCV
• numpy
• scipy
• easydict

Installation

git clone https://github.com/shepnerd/outpainting_srn.git
cd outpainting_srn/

Testing

• Download the pretrained models through the following links (CelebA-HQ_256, Paris streetview, Cityscapes), and unzip and put them into checkpoints/.

• To test images in a folder, we call test.py with the opinion --dataset_path and --load_model_dir. We give some CelebA-HQ_256 examples in the imgs/celebahq-256/. For example:

  python test.py --dataset celebahq-256 --data_file ./imgs/celebahq-256/ --load_model_dir ./checkpoints/celebahq-srn-subpixel --random_mask 1
  

  or write / modify test.sh according to your own needs, then execute this script as (Linux platform):

  sh ./test.sh
  

The visual evaluations will be saved in the folder ./test_results/.

Training

• Training the model with the reconstruction loss (relative spatial variant loss) firstly (set the opinion --pretrain_network 1), then fine-tuning the model with all losses (--pretrain_network 0 and --load_model_dir [Your model path]) after the first stage converges. To pretrain the network,

  python train.py --dataset [DATASET_NAME] --data_file [DATASET_TRAININGFILE] --gpu_ids [NUM] --pretrain_network 1 --batch_size 16
  

  where [DATASET_TRAININGFILE] indicates a file storing the full paths of the training images. A simple example is given as:

  python train.py --dataset celebahq-256 --data_file ../celebahq-256_train.txt --gpu_ids 0 --img_shapes 256,256 --pretrain_network 1 --batch_size 8
  

• Then finetune the network,

  python train.py --dataset [DATASET_NAME] --data_file [DATASET_TRAININGFILE] --gpu_ids [NUM] --pretrain_network 0 --load_model_dir [PRETRAINED_MODEL_PATH] --batch_size 8
  

Datasets

All used datasets (CelebA-HQ, CUB200, Dog, DeepFashion, Paris-Streetview, Cityscape, and Places2) and their corresponding train/test splits are given in the paper.

Todo

• Other pretrained models
• ...

Disclaimer

• The training and evaluation performance on large-scale datasets (with a variety of categories) is unstable due to possible mode collapse in adversarial training.
• The used ID-MRF is a simplified version based on contextual loss. The step of excluding s is omitted for computational efficiency.

Citation

If our method is useful for your research, please consider citing:

@inproceedings{wang2019srn,
  title={Wide-Context Semantic Image Extrapolation},
  author={Wang, Yi and Tao, Xin and Shen, Xiaoyong and Jia, Jiaya},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={1399--1408},
  year={2019}
}

@inproceedings{wang2018image,
  title={Image Inpainting via Generative Multi-column Convolutional Neural Networks},
  author={Wang, Yi and Tao, Xin and Qi, Xiaojuan and Shen, Xiaoyong and Jia, Jiaya},
  booktitle={Advances in Neural Information Processing Systems},
  pages={331--340},
  year={2018}
}

Acknowledgments

Our code is built upon Image Inpainting via Generative Multi-column Convolutional Neural Networks and pix2pixHD.

Recent related work

Contact

Please send email to [email protected].