Automatic Image Colorization: a comparative overview

Project for the university course of Vision and Cognitive Services.

We used the following datasets:

• Imagenette: https://github.com/fastai/imagenette
• Pascal: https://deepai.org/dataset/pascal-voc
• Places: https://paperswithcode.com/dataset/places205
• Birds: https://www.kaggle.com/gpiosenka/100-bird-species
• Flowers: https://www.robots.ox.ac.uk/~vgg/data/flowers/102/

Here you can find the pretrained models used in this project:

• Dahl: https://tinyclouds.org/colorize/ (Download section)
• Zhang eccv_16: https://github.com/richzhang/colorization/blob/master/colorizers/eccv16.py - related paper here
• Zhang siggraph17: https://github.com/richzhang/colorization/blob/master/colorizers/siggraph17.py - related paper here
• ChromaGAN: https://github.com/pvitoria/ChromaGAN - related paper here
• InstColorization: https://github.com/ericsujw/InstColorization - related paper here

Requirements

• python 3.6 or 3.8
• virtualenv wrapper: https://virtualenvwrapper.readthedocs.io/en/latest/

Dahl

• download the pretrained model and place the colorize.tfmodel file in the models folder

• create a virtual environment: mkvirtualenv --python=python3 dahl

• install the requirements: pip install -r requirements_dahl.txt

• position yourself into the following folder: cd src/models

• run the model: python3 dahl.py

Eccv16 and Siggraph17

• create a virtual environment: mkvirtualenv --python=python3 zhang

• install the requirements: pip install -r requirements_zhang.txt

• position yourself into the following folder: cd src/models

• run the model: python3 Eccv16andSiggraph17.py

ChromaGAN

• download the pretrained model and place the ChromaGAN.h5 file in the models folder

• create a virtual environment using python 3.6: mkvirtualenv --python=python3 chromaGAN

• install the requirements: pip install -r requirements_chromaGAN.txt

• position yourself into the following folder: cd src/models

• run the model: python3 chromaGAN.py

InstColorization

The code to run this model is contained in the following notebook: src/models/InstColorization.ipynb.

Metrics

• AlexNet classification accuracy
• Turing test
• LPIPS: https://github.com/richzhang/PerceptualSimilarity
• PSNR and SSIM: https://cvnote.ddlee.cc/2019/09/12/psnr-ssim-python