Click Here to read the full experiment of DC-GAN
I downloaded thousands of pictures of my favorite anime character, Chitanda Eru, in order to build a face dataset for my upcoming GAN tests. Building a dataset is not a easy thing, since the pictures which are downloaded and processed by some automatic program are also unqualified for precise training. My goal is to generate one person but not all kinds of anime characters, so the process cannot be a simple selection for anime pictures but need to choice the right-positioned face and eliminate incompletion and fuzziness.
In order to optimize pictures according to standard mentioned below, I tried to adapt a headshot cropper using OpenCV to detect anime headshots in dataset. With adjustment of some cropping parameters, the result seemed very nice. However, when I briefly scanned the cropped pictures, I found a lot of collapses and other characters in anime Hyouka. It was no time for me to train another classifier to check these pictures, so I tried to quickly scan the dataset and manually delete the items did not reach the standard.
I will create a larger set and upload them when the whole project complete. Thanks for your attention.