/jp/ - Otaku Culture

>>23892506
So, I tried this library a few weeks ago, but I couldn't be bothered to write up a post about it. I guess I better do it now before the thread dies?
So, lbpcascade_animeface is pretty easy to install and use, but there are a lot of faces it won't detect. You could use it to auto-crop a few images, but you'd still have to do a lot of manual work.
The author has also written another, more sophisticated program for detecting anime faces, at:
https://github.com/nagadomi/animeface-2009
I needed to install a ruby-dev package and some imagemagick related things, but now it runs. This script detect a lot more faces, including a sizeable amount of false positives. I'd still say it's fairly suitable for creating collages like this, since it is easier to delete false positives than it is to crop images.

Both libraries do poorly on black and white images, and lineart - at least for the image set I tried it on. It might be possible to circumvent this by generating black and white and lineart version of each image in the training dataset (the ruby script is pretrained, but the documentation tells you how to train it on your own dataset) and train on this extended dataset. I'm probably just going to stick to using the default model, though, as I don't need to use BW images and doodles for my collage.

Attached pic is an output of the sample script for the ruby program, illustrating a false positive. Wrinkles/creases on clothing is another thing that gives me a lot of false positives.