The first sentence sets the mood, it its particularly delightful to have a paper start of with a strong sentence. Zhang et. al knock it out of the park by getting you to ponder only 6 words in. I’m excited about finishing this paper.
This paper addresses the problem of colorizing greyscale images. It is important to point out that the paper is not trying to create an accurate model that can reproduce the image. That is near impossible because you have already lost two out of your three dimensions (color and depth). However, it is possible to create a plausible colorization that is believable. That is the aim of the model; however the paper aims a little higher.
This paper claims three contributions:
Progress on Colorization particularly in three areas:
- Designing an objective function able to cope with the multimodal reality of recolorization
- Introducing a novel way of testing algorithms
- Setting a new high water mark
Self-Supervised Benchmarking They note their addition of the colorization task as a competitive metric useful for evaluating algorithms following self-supervised representation learning.
Tomorrow, I’ll probably look at the model in more detail: