Open cpufreak opened 6 years ago
Cool, thanks for sharing!
Part of the problem is resolution. The method scales the image down to 256x256, and outputs colors at 64x64 resolution. The faces are very small and easy to miss.
More modern methods (eg https://arxiv.org/abs/1705.07208) can also capture correlation between pixels as well.
so I found the Prokudin-Gorskii photo collection of early color prints, and to a benefit, the collection includes the original B+W negatives, which allowed for easy comparison. some images such as this (the original color print for comparison here ) turn out amazingly well. However, to contrast that, some of them turn out quite bad such as this one, especially compared to the digitally retouched version of the original. is this perhaps just an issue of needing more training data?