A couple years ago I used deep learning to recolor some old family photos. The net I was using would do it's best guess first, then present a limited set of opt

A couple years ago I used deep learning to recolor some old family photos. The net I was using would do it's best guess first, then present a limited set of options for me to choose from when it didn't know what to do. In general, its first pass would immediately make human faces look correct, but leave everything else really desaturated. I think some of it comes down to the fact that convnets are really good at recognizing human "shapes" (faces, silhouettes, etc) and also the valid color space for human skin tone is pretty constrained if the photo was well lit and you have the luminance value for a grayscale image. The network was particularly bad at getting the colors of balloons in a picture. I don't know what the real colors were, but the network was trying to tell me 4 of the balloons in a picture were pale yellow and beige while one was a super saturated red.