Skip to content

Latest commit

 

History

History
112 lines (84 loc) · 5.8 KB

README.md

File metadata and controls

112 lines (84 loc) · 5.8 KB
Raw

Deep Exemplar-based Colorization

This is the implementation of paper Deep Exemplar-based Colorization by Mingming He*, Dongdong Chen*, Jing Liao, Pedro V. Sander and Lu Yuan in ACM Transactions on Graphics (SIGGRAPH 2018) (*indicates equal contribution).

Introduction

Deep Exemplar-based Colorization is the first deep learning approach for exemplar-based local colorization. Given a reference color image, our convolutional neural network directly maps a grayscale image to an output colorized image.

image

The proposed network consists of two sub-networks, Similarity Sub-net which computes the semantic similarities between the reference and the target, and Colorization Sub-net which selects, propagates and predicts the chrominances channels of the target.

The input includes a grayscale target image, a color reference image and bidirectional mapping functions. We use Deep Image Analogy as default to generate bidirectional mapping functions. It is applicable to replace with other dense correspondence estimation algorithms.

The code of the part Color Reference Recommendation is now released. Please refere to Gray-Image-Retrieval for more details.

For more results, please refer to our Supplementary.

(Update) If you would to compile on Linux, please try this repository: https://github.com/ncianeo/Deep-Exemplar-based-Colorization/tree/linux-docker-cv-caffe-build, thank ncianeo for solving this issue.

License

© Microsoft, 2017. Licensed under a MIT license.

Getting Started

Prerequisites

  • Similarity Sub-net:

    • Windows (64bit)
    • NVIDIA GPU (CUDA 8.0 & CuDNN 5)
    • Visual Studio 2013

  • Colorization Sub-net:

    • Pytorch & the 3rd party Python libraries (OpenCV, scikit-learn and scikit-image)

Build

Similarity Sub-net is implemented in C++ combined with CUDA and requires compiling in Visual Studio as follows:

  • Build Caffe at first. Just follow the tutorial here.
  • Edit similarity_combo.vcxproj under similarity_subnet\windows\similarity_combo\ to make the CUDA version in it match yours.
  • Open solution Caffe and add similarity_combo.vcxproj.
  • Build project similarity_combo.
  • (Optional) If you use Deep Image Analogy, please add deep_image_analogy.vcxproj under similarity_subnet\windows\deep_image_analogy\ and build it.

Download Models

You need to download models before running a demo.

Demo

We prepare an example under the folder demo\ with:

(1) Input data folder example\ including two parts:

  • A folder input\ with the input images (grayscale target images and color reference images) inside.
  • A file pairs.txt to specify a target, a reference and a flag (1 as default) as an example in each line, e.g., 
    in1.jpg ref1.jpg 1
in2.jpg ref2.jpg 1
...

(2) Executable script run.bat including three commands:

  • (Optional) A command to generate bidirectional mapping functions using Deep Image Analogy:

    deep_image_analogy.exe [MODEL_DIR] [INPUT_ROOT_DIR] [START_LINE_ID] [END_LINE_ID] [GPU_ID]
e.g., exe\deep_image_analogy.exe models\deep_image_analogy\ example\ 0 2 0

    (Note if you use other algorithms for bidirectional mapping functions, please generate flow files referring to the format of those by Deep Image Analogy and put them to the folder example\flow\.)

  • A command to generate similarity maps for colorization (Similarity Subnet):

    similarity_combo.exe [MODEL_DIR] [INPUT_ROOT_DIR] [START_LINE_ID] [END_LINE_ID] [GPU_ID]
e.g., exe\similarity_combo.exe models\similarity_subnet\ example\ 0 2 0

  • A command to do colorization with our pretrained model (Colorization Subnet):

    python test.py --short_size [SHORT_EDGE_SIZE] --test_model [MODEL_FILE] --data_root [INPUT_ROOT_DIR] --out_dir [OUTPUT_DIR] --gpu_id [GPU_ID]
e.g., python ..\colorization_subnet\test.py --short_size 256 --test_model models\colorization_subnet\example_net.pth --data_root example\ --out_dir example\res\ --gpu_id 0

Run

We provide pre-built executable files in folder demo\exe\, please try them.

Tips

Our test input images are resized to w x h (min(w, h)=256) considering the cost of computing bidirectional mapping functions by Deep Image Analogy. But we also support higher resolution input images.

Citation

If you find Deep Exemplar-based Colorization helpful for your research, please consider citing:

@article{he2018deep,
  title={Deep exemplar-based colorization},
  author={He, Mingming and Chen, Dongdong and Liao, Jing and Sander, Pedro V and Yuan, Lu},
  journal={ACM Transactions on Graphics (TOG)},
  volume={37},
  number={4},
  pages={47},
  year={2018},
  publisher={ACM}
}