Comment 9 for bug 1582885

Ah, I remember there being announcements about vrty.org. I should have made the connection. I re-read the March 2017 post just now.

When I first posted this, I imagined a simpler person-detector that just filtered *all* images with people in them, rather than making a NSFW judgement. But maybe the yahoo model that Brad mentions would work, though it would require Caffe, which I imagine is a difficult-to-guarantee requirement. Since I think that the cpu-only version of tensorflow would be pretty easy to require (installs in a Python 3 environment with pip only), I tried running it through caffe-tensorflow[1], but that failed. However, if it's just a person filter that we want, maybe the yahoo model wouldn't be needed. And, since speed isn't an issue, maybe it could be deployed with Numpy only.

With a few thousand labeled images, it might be possible to use, say, a small pre-trained VGG network to do this. Peter, does such data exist? Grepping through the source, I see only PIL and cairo as, I think, non-stdlib imports. CouldnNumpy or tensorflow (not tensorflow-gpu) be added?

[1]: https://github.com/ethereon/caffe-tensorflow