Colorful Image Colorization

Implementation of Colorful Image Colorization by Richard Zhang.

Paper - Code - YouTube - Website

Main ideas

Objective Function

Not the Euclidean Loss, but Multinominal Cross Entropy Loss.

Class Rebalancing

Observe that the number of pixels in natural images at desaturated values are orders of magnitude higher than for saturated values. Therefore, the loss function is dominated by desaturated ab values.

Class Probabilities to Point Estimates

Not mode of the predicted distribution for each pixel, not its mean, but its annealed-mean.

Reference

• https://github.com/sizhky/Colorization
• https://github.com/nku-shengzheliu/Pytorch-Colorful-Image-Colorization
• https://github.com/87003697/colorful_image_colorization_pytorch
• https://github.com/yasamanhbn/Image-Colorization
• https://github.com/luxiya01/deep-coloring
• https://github.com/dtpreda/colorful
• https://github.com/stanleynguyen/corolization