Let's say you have too many images (extracted from many videos) — like, 100,000, and you want to use a smart automated process to reduce that to 10,000, which is the maximum recommended number for training in the FAQ.
Would it make sense to use the identity information in the alignments file — along with one of the sorting algorithms provided by Faceswap — to:
- first order the images
- then start deleting images with minimum distance (difference) to its neighbor until you've hit your goal (100k -> 10k)
In order to end up with a set with maximum variety?
Lastly, is there a document with definitions for each of the many sorting algorithms?
(none,blur,blur-fft,distance,face,face-cnn,face-cnn-dissim,yaw,pitch,roll,hist,hist-dissim,color-black,color-gray,color-luma,color-green,color-orange,size)