Hacker News new | comments | ask | show | jobs | submit login

author here: let me know if you have any questions

How well does your solution handle non-white-people NSFW images?

I know it's only a matter of training with other inputs, but most solutions I've seen haven't been trained with enough non-Caucasian images to give good results. Since this at least seems to be detecting shape regardless of skin tone it has more potential, but it'll only be as good as its training set.

It was trained on non-white-people NSFW images to (eg. non-white people, hentai). However, I don't have per genre numbers right now.

I'm sure it's easy to find "training data" for this project, but FYI 4walled dot cc is probably a GREAT source. It's an archive of thousands of 4chan images, and most are already classified by SFW / Borderline / NSFW. Since it came from 4chan, it includes lots of non-pornographic NSFW things as well (gore, swastikas, etc.), and lots of duplicates with various cropping, reencodimg and other processing done to them.

Is the idea that people would submit all images to your pre trained API / program, or would they also be generating a training set themselves? How sensitive is the algorithm to incorrectly tagged images in the training set?

Right now we've got the fully-trained convnet deployed here: https://developer.clarifai.com/guide/tag#nsfw

I have not quantified how robust this net is in the face of noisy labels.

How well does it work on grayscale images?

EDIT: Just created an account and tried it out on a test image, and it returned "nsfw" with prob 0.998, which is pretty amazing.

EDIT2: It's much less strong against oblique views, even with full color. http://40.media.tumblr.com/cfeea64370e3b1228906472ebd8f344e/... (NSFW, obviously!) only returns 0.541.

Awesome, fwiw you can hit the demo here without getting an API key http://clarifai.com/?model=nsfw-v1.0&probs=1

Could you do a deep dream run and generate porn images from the training set?

I don't think the result could be called "porn" in a productive sense for anyone other than the Great Old Ones.

How did you come up with 64x64 sliding windows?

Just tried a few heat maps and that one looked good.

Why did you have no male examples?

Look closer at the last figure.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact