Which tells me it’s almost certainly something like this short little Python snippet:
ycbcr = im.convert("YCbCr")
pixels = ycbcr.getdata()
p = [x for x in pixels if 80 <= x <= 120 and 133 <= x <= 173]
This returns a list of the pixels that fall within a certain gamut. The idea here is that if you compare the length of this list to the original image, you get a percentage of pixels that are flesh-colored, and it works relatively well for just about any skintone because it only checks the color channels and ignores luminance. It also works for cream-colored curtains, adobe, and several other things like you see here. It’s quick to implement and takes a lot less effort and computational power than a machine learning algorithm.
That gamut comes from a paper that may or may not be patent-encumbered, I don’t know.
Anyway, my guess is that they’ve got this, and maybe an OpenCV-based nipple detector.
This is allegedly their panicked response to being overrun by porn bots, some of them posting kiddie porn. Considering how much of their traffic was actual porn, I’m guessing this is probably going to kill Tumblr.
I could not agree more. The people who cry about being “silenced” when their crappy alt-right views get removed by YouTube, not because FOSTA SESTA holds content hosting responsible for prostitution ads, but because it holds content hosting responsible, period. I mean, if I wanted to take someone down in a cruel way and found out they had an insecure business website, I could deface their website with a bunch of “barely 18” content in the middle of the night with fake solicitation contact info and contact the authorities if we continue down this road of punishing the hosting party. It’s doubtful law enforcement would ever hold a party responsible for defacing…but we live in perilously stupid times.