knazeri/edge-connect

EdgeConnect: Structure Guided Image Inpainting using Edge Prediction, ICCVW 2019 https://arxiv.org/abs/1901.00212

EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning

ArXiv | BibTex

Introduction:

We develop a new approach for image inpainting that does a better job of reproducing filled regions exhibiting fine details inspired by our understanding of how artists work: lines first, color next. We propose a two-stage adversarial model EdgeConnect that comprises of an edge generator followed by an image completion network. The edge generator hallucinates edges of the missing region (both regular and irregular) of the image, and the image completion network fills in the missing regions using hallucinated edges as a priori. Detailed description of the system can be found in our paper.

Prerequisites

• Python 3
• PyTorch 1.0
• NVIDIA GPU + CUDA cuDNN

Installation

• Clone this repo:

git clone https://github.com/knazeri/edge-connect.git
cd edge-connect

• Install PyTorch and dependencies from http://pytorch.org
• Install python requirements:

pip install -r requirements.txt

Datasets

1) Images

We use Places2, CelebA and Paris Street-View datasets. To train a model on the full dataset, download datasets from official websites.

After downloading, run scripts/flist.py to generate train, test and validation set file lists. For example, to generate the training set file list on Places2 dataset run:

mkdir datasets
python ./scripts/flist.py --path path_to_places2_train_set --output ./datasets/places_train.flist

2) Irregular Masks

Our model is trained on the irregular mask dataset provided by Liu et al.. You can download publically available Irregular Mask Dataset from their website.

Alternatively, you can download Quick Draw Irregular Mask Dataset by Karim Iskakov which is combination of 50 million strokes drawn by human hand.

Please use scripts/flist.py to generate train, test and validation set masks file lists as explained above.

Getting Started

Download the pre-trained models using the following links and copy them under ./checkpoints directory.

Places2 | CelebA | Paris-StreetView

Alternatively, you can run the following script to automatically download the pre-trained models:

bash ./scripts/download_model.sh

1) Training

To train the model, create a config.yaml file similar to the example config file and copy it under your checkpoints directory. Read the configuration guide for more information on model configuration.

EdgeConnect is trained in three stages: 1) training the edge model, 2) training the inpaint model and 3) training the joint model. To train the model:

python train.py --model [stage] --checkpoints [path to checkpoints]

For example to train the edge model on Places2 dataset under ./checkpoints/places2 directory:

python train.py --model 1 --checkpoints ./checkpoints/places2

Convergence of the model differs from dataset to dataset. For example Places2 dataset converges in one of two epochs, while smaller datasets like CelebA require almost 40 epochs to converge. You can set the number of training iterations by changing MAX_ITERS value in the configuration file.

2) Testing

To test the model, create a config.yaml file similar to the example config file and copy it under your checkpoints directory. Read the configuration guide for more information on model configuration.

You can test the model on all three stages: 1) edge model, 2) inpaint model and 3) joint model. In each case, you need to provide an input image (image with a mask) and a grayscale mask file. Please make sure that the mask file covers the entire mask region in the input image. To test the model:

python test.py \
  --model [stage] \
  --checkpoints [path to checkpoints] \
  --input [path to input directory or file] \
  --mask [path to masks directory or mask file] \
  --output [path to the output directory]

We provide some test examples under ./examples directory. Please download the pre-trained models and run:

python test.py \
  --checkpoints ./checkpoints/places2 
  --input ./examples/places2/images 
  --mask ./examples/places2/masks
  --output ./checkpoints/results

This script will inpaint all images in ./examples/places2/images using their corresponding masks in ./examples/places2/mask directory and saves the results in ./checkpoints/results directory. By default test.py script is run on stage 3 (--model=3).

3) Evaluating

To evaluate the model, you need to first run the model in test mode against your validation set and save the results on disk. We provide a utility ./scripts/metrics.py to evaluate the model using PSNR, SSIM and Mean Absolute Error:

python ./scripts/metrics.py --data-path [path to validation set] --output-path [path to model output]

To measure the Fréchet Inception Distance (FID score) run ./scripts/fid_score.py. We utilize the PyTorch implementation of FID from here which uses the pretrained weights from PyTorch’s Inception model.

python ./scripts/fid_score.py --path [path to validation, path to model output] --gpu [GPU id to use]

Alternative Edge Detection

By default, we use Canny edge detector to extract edge information from the input images. If you want to train the model with an external edge detection (Holistically-Nested Edge Detection for example), you need to generate edge maps for the entire training/test sets as a pre-processing and their corresponding file lists using scripts/flist.py as explained above. Please make sure the file names and directory structure match your training/test sets. You can switch to external edge detection by specifying EDGE=2 in the config file.

Model Configuration

The model configuration is stored in a config.yaml file under your checkpoints directory. The following tables provide the documentation for all the options available in the configuration file:

General Model Configurations

Loading Train, Test and Validation Sets Configurations

Training Mode Configurations

License

Licensed under a Creative Commons Attribution-NonCommercial 4.0 International.

Except where otherwise noted, this content is published under a CC BY-NC license, which means that you can copy, remix, transform and build upon the content as long as you do not use the material for commercial purposes and give appropriate credit and provide a link to the license.

Citation

If you use this code for your research, please cite our paper EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning:

@inproceedings{nazeri2019edgeconnect,
  title={EdgeConnect: Generative Image Inpainting with Adversarial Edge Learning},
  author={Nazeri, Kamyar and Ng, Eric and Joseph, Tony and Qureshi, Faisal and Ebrahimi, Mehran},
  journal={arXiv preprint},
  year={2019},
}