/Reference_based_Skectch_Image_Colorization

This repository implements the paper "Reference based Sketch Image Colorization using Augmented-Self Reference and Dense Semantic Correspondence" which is published in CVPR2020.

Primary LanguagePython

Reference_based_Skectch_Image_Colorization

This repository implements the paper "Reference based Sketch Image Colorization using Augmented-Self Reference and Dense Semantic Correspondence" published in CVPR2020.

Requirements

  • python3.6+
  • pytorch 1.6.0
  • others.

Usage

training a model

python3 main.py --config config.yml

testing a model

Not implmented yet

Architecture

architecture

Results

shifted_result normalresult1 normalresult2

Comments

In this implementation, the triplet loss function is meaningless. It always show zeros for scaled dot product and l2 norm distance, if I am wrong, please make issue. Without the triplet loss, we can obtain good results. Even if a model is trained only 2 epochs, the model shows meaningful results.

Reference

  1. tps_transform : https://github.com/cheind/py-thin-plate-spline
  2. spectral normalization : https://github.com/christiancosgrove/pytorch-spectral-normalization-gan/blob/master/spectral_normalization.py
  3. unet : https://github.com/milesial/Pytorch-UNet
  4. dataset : https://www.kaggle.com/ktaebum/anime-sketch-colorization-pair