speeq 8 days ago

It would be nice if someone could compare these commercially available APIs with Yahoo's open_nsfw model in terms of accuracy: https://github.com/yahoo/open_nsfw

I'm currently building an API wrapper around it and running it on a Hetzner server with a GTX 1080 - prediction takes about 0.25 seconds and while I haven't optimised it for parallel execution, I think it should be able to handle at least +10 images/sec comfortably. I'm also testing video moderation by using ffmpeg to slice the video into screenshots and predicting the min/avg/high scores.

Moderating 25 million explicit images using Google Cloud Vision would cost around $19,500/mo vs €99/mo on Hetzner.

  • mohi13 8 days ago

    Makes a lot of sense, actually its really difficult to get a large enough dataset for moderation tasks to make a decent inhouse model for a fair enough comparison.

    Sure, we can try scraping that from pornhub etc but fee then the negative classes would be very domain specific, using stock images may not provide a good measure.

    Also, its really weird to assign such a task to any of your employees, feels kinda strange :)

    • speeq 8 days ago

      Yahoo's model could be fine-tuned: http://caffe.berkeleyvision.org/gathered/examples/finetune_f...

      Yeah, it's definitely not a nice task but what's stopping someone (well, besides potential legal issues) from using these commercial APIs to create datasets programatically and training a cloned model from that?

      I'm curious what the profit margins are on these APIs because I think they are way overpriced.

konradzikusek 8 days ago

I really hope you haven’t picked a dataset that features only women nudity, but sample images suggest otherwise: https://dataturks.com/projects/Mohan/NSFW(Nudity%20Detection...

yaseen-rob 8 days ago

I tried out various image labeling APIs, including Google Vision (Safe Search) for exactly this use case (moderation). I was honestly astonished at the pricing of these APIs. Google is somewhere at 1.50€ for 1000 images which is - imo - very expensive. I tried out the default models that come with Tensorflow but well, they are trained on scientific datasets which typically involve species and flowers - no luck there either. Any good tips for pre-trained models that solve this (for tensorflow)?

  • dannyw 8 days ago

    Download 1000 NSFW images, 1000 racy images and 100 completely SFW images. Train a model and publish it on GitHub?

    • mohi13 8 days ago

      Would be really interested to see the results of this.

      Also, consider using Dataturks to create and host the dataset.

jaequery 8 days ago

what is the verdict? i been just scrolling down to see which is best and cant find any.

edent 8 days ago

The sample images shown only feature white women. Is that a limitation of the dataset?

  • specializeded 8 days ago

    Tila Tequila is Vietnamese, and features in the majority of the sample images shown.

    • drb91 8 days ago

      At the same time, the data set is atrocious.

    • nahom1 8 days ago

      Didn't know she was a famous model :)

    • ummonk 8 days ago

      She's also a white nationalist and neo-Nazi...

mohi13 8 days ago

BTW we had a question in general to anyone who can help us, Does hosting such a dataset cause issues with SEO etc? Anything else we should be aware of?

  • judge2020 8 days ago

    Probably wouldn't be wise to host a dataset like this with "example" images embedded or linked where a search engine could find them.

    • mohi13 8 days ago

      thanks, will see how can I block crawl on this dataset. BTW how does it hurt exactly, couldn't find much except in case of safe-search mode.