Globally and Locally Consistent Image Completion

Satoshi Iizuka, Edgar Simo-Serra, Hiroshi Ishikawa

News

09/17/2020 Update: We have released the following two models:

• completionnet_places2_freeform.t7: An image completion model trained with free-form holes on the Places2 dataset, which will work better than the model trained with rectangular holes, even without post-processing. We used a part of the context encoder [Pathak et al. 2016] implementation to generate the random free-form holes for training.
• completionnet_celeba.t7: A face completion model trained with rectangular holes on the CelebA dataset. This model was trained on face images with the smallest edges in the [160, 178], and thus it will work best on images of similar sizes.

These models can be downloaded via download_model.sh.

Overview

This code provides an implementation of the research paper:

  "Globally and Locally Consistent Image Completion"
  Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa
  ACM Transaction on Graphics (Proc. of SIGGRAPH 2017), 2017

We learn to inpaint missing regions with a deep convolutional network. Our network completes images of arbitrary resolutions by filling in missing regions of any shape. We use global and local context discriminators to train the completion network to provide both locally and globally consistent results. See our project page for more detailed information.

License

  Copyright (C) <2017> <Satoshi Iizuka, Edgar Simo-Serra, Hiroshi Ishikawa>

  This work is licensed under the Creative Commons
  Attribution-NonCommercial-ShareAlike 4.0 International License. To view a copy
  of this license, visit http://creativecommons.org/licenses/by-nc-sa/4.0/ or
  send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.

  Satoshi Iizuka, University of Tsukuba
  [email protected], http://iizuka.cs.tsukuba.ac.jp/index_eng.html
  
  Edgar Simo-Serra, Waseda University
  [email protected], https://esslab.jp/~ess/

Dependencies

• Torch7
• nn
• image
• nngraph
• torch-opencv (optional, required for post-processing)

The packages of nn, image, and nngraph should be a part of a standard Torch7 install. For information on how to install Torch7 please see the official torch documentation on the subject. The torch-opencv is OpenCV bindings for LuaJit+Torch, which can be installed via luarocks install cv after installing OpenCV 3.1. Please see the instruction page for more detailed information.

17/09/2020 Note: If you fail to install Torch7 with current GPUs, please try the self-contained Torch installation repository (unofficial) by @nadadomi, which supports CUDA10.1, Volta, and Turing.

Usage

First, download the models by running the download script:

bash download_model.sh

Basic usage is:

th inpaint.lua --input <input_image> --mask <mask_image>

The mask is a binary image (1 for pixels to be completed, 0 otherwise) and should be the same size as the input image. If the mask is not specified, a mask with randomly generated holes will be used.

Other options:

• --model: Model to be used. Defaults to 'completionnet_places2_freeform.t7'.
• --gpu: Use GPU for the computation. cunn is required to use this option. Defaults to false.
• --maxdim: Long edge dimension of the input image. Defaults to 600.
• --postproc: Perform the post-processing. Defaults to false. If you fail to install the torch-opencv, do not use this option to avoid using the package.

For example:

th inpaint.lua --input example.png --mask example_mask.png

Best Performance

• The Places models were trained on the Places2 dataset and thus best performance is for natural outdoor images.
• While the Places2 models work on images of any size with arbitrary holes, we trained them on images with the smallest edges in the [256, 384] pixel range and random holes in the [96, 128] pixel range. Our models will work best on images with holes of those sizes.
• Significantly large holes or extrapolation when the holes are at the border of images may fail to be filled in due to limited spatial support of the model.

Notes

• This is developed on a Linux machine running Ubuntu 16.04 during late 2016.
• Provided model and sample code is under a non-commercial creative commons license.

Citing

If you use this code please cite:

@Article{IizukaSIGGRAPH2017,
  author = {Satoshi Iizuka and Edgar Simo-Serra and Hiroshi Ishikawa},
  title = {{Globally and Locally Consistent Image Completion}},
  journal = "ACM Transactions on Graphics (Proc. of SIGGRAPH)",
  year = 2017,
  volume = 36,
  number = 4,
  pages = 107:1--107:14,
  articleno = 107,
}