AI detection of headscissors images

[wdrocked]

Is anyone here even remotely in touch with ai ways and methods of training it to detect headscissors pictures and detect them? Would it be feasable just to learn some basic ai shit and just throw pircures at it, or is there more work involved?
I want to search through some parts of the internet for such pictures but theres a large part of them which dont have any text description alluding to headscissors, requiring manual review or ai detection.

[uxiel]

There's probably a relatively small amount of data for CV systems to learn from.

[Only Registered Users Can See LinksClick Here To Register]

Computer Vision: For image classification using deep learning, a rule of thumb is 1,000 images per class, where this number can go down significantly if one uses pre-trained models

The real work would likely be getting data and putting it in the right format, not building systems. As the saying goes, data scientists are mostly just data janitors.

[wdrocked]

Originally Posted by uxiel [Only Registered Users Can See LinksClick Here To Register]

There's probably a relatively small amount of data for CV systems to learn from.

[Only Registered Users Can See LinksClick Here To Register]



The real work would likely be getting data and putting it in the right format, not building systems. As the saying goes, data scientists are mostly just data janitors.

so lets say we have only one class, headscissors i guess, 1000 pics? yeah there are 1000 pics in these forums easy. work required? well yeah but maybe some of us forum members could collaborate on making this 1k database. the non headscissors nude-erotic ones (the false pictures for the ai to not detect) could just be mined from some random porn sites, because any random erotic picture has like a 0.001% chance of being headscissors

[uxiel]

I guess the classes would be "headscissors" and "not headscissors." However, for the "not" category, you might want to carefully select pictures where the context is at least similar, like a man and woman sitting next to each other, or a chokehold, etc., and not just random pictures of apples and birds. In many real systems, the "no" class will make up 99%+ of the data, so you might actually need a lot more data for non-headscissors. I'd personally lean towards collecting *a lot* of "no" data.

If you're building out a system, I'd also recommend first building out something to differentiate cars vs. busses or apples vs. oranges, so you're at least confident you know how to build (really, use) the system/model properly. That way, if something doesn't work, you have some idea if it's a bad implementation/system hyperparameters/etc., or if it's just bad/insufficient data. I've done ML before, but not CV specifically, so I can't tell you of specific gotchas. I'm not sure if their "per class" means yes/no, or the X/Y/Z in X / Y / Z / none of the above.

Anyway, it would be a lot of work, and if you wanted to recruit helpers, you might have to convince them of the value (example: "I'm building a web crawler and feeding all the images into this classifier to find every headscissors picture on the planet" -- and in that case, you'd probably want to make this system fairly generic so you can easily find all the pics of chokeholds, or feet, or whatever other people are into so that they don't have to reinvent the wheel every time), or if it's just something you find neat.

[wdrocked]

Originally Posted by uxiel [Only Registered Users Can See LinksClick Here To Register]

I guess the classes would be "headscissors" and "not headscissors." However, for the "not" category, you might want to carefully select pictures where the context is at least similar, like a man and woman sitting next to each other, or a chokehold, etc., and not just random pictures of apples and birds. In many real systems, the "no" class will make up 99%+ of the data, so you might actually need a lot more data for non-headscissors. I'd personally lean towards collecting *a lot* of "no" data.

If you're building out a system, I'd also recommend first building out something to differentiate cars vs. busses or apples vs. oranges, so you're at least confident you know how to build (really, use) the system/model properly. That way, if something doesn't work, you have some idea if it's a bad implementation/system hyperparameters/etc., or if it's just bad/insufficient data. I've done ML before, but not CV specifically, so I can't tell you of specific gotchas. I'm not sure if their "per class" means yes/no, or the X/Y/Z in X / Y / Z / none of the above.

Anyway, it would be a lot of work, and if you wanted to recruit helpers, you might have to convince them of the value (example: "I'm building a web crawler and feeding all the images into this classifier to find every headscissors picture on the planet" -- and in that case, you'd probably want to make this system fairly generic so you can easily find all the pics of chokeholds, or feet, or whatever other people are into so that they don't have to reinvent the wheel every time), or if it's just something you find neat.

its just sonething i find neat, and i also should crawl through some instagram profiles
im putting this thread here if anyone could actually do it, but nevertheless ill probably do whats needed and get into ml for this specific reason
anyways i think it would maybe maybe be easier if an already made ai could detect peoples poses in 3d and then i could use that to detect if a head is between thighs

[uxiel]

For IG, the crawling would be the first interesting challenge IMO. The way the page is structured, the image is inside a bunch of nested divs, and would take more than 2 minutes to write a generic scraper for (instagram would probably also rate limit you if you start blasting them with traffic). Of course you could manually grab pictures 1x1 by going to your browser's element inspector, but it'd take some time. There are of course websites that will let you paste in a post URL and give you the image directly, but they don't generally support bulk operations from what I've seen.

Anyway, I'd probably first play around with basic classifiers on simple things (identifying fruits or whatever you see in online tutorials), then scrape for the data you need, then play around with building the classifier you actually want.

[InTheFlesh86]

If everyone could go ahead and send me 1,000 of their best headscissors pics I can start working on...uh...synthesizing that data right away.

[uxiel]

Originally Posted by InTheFlesh86 [Only Registered Users Can See LinksClick Here To Register]

If everyone could go ahead and send me 1,000 of their best headscissors pics I can start working on...uh...synthesizing that data right away.

I will send you 1000 pics of not-headscissors.

[wdrocked]

Originally Posted by InTheFlesh86 [Only Registered Users Can See LinksClick Here To Register]

If everyone could go ahead and send me 1,000 of their best headscissors pics I can start working on...uh...synthesizing that data right away.

just trust me bro

Originally Posted by wdrocked [Only Registered Users Can See LinksClick Here To Register]

Is anyone here even remotely in touch with ai ways and methods of training it to detect headscissors pictures and detect them? Would it be feasable just to learn some basic ai shit and just throw pircures at it, or is there more work involved?
I want to search through some parts of the internet for such pictures but theres a large part of them which dont have any text description alluding to headscissors, requiring manual review or ai detection.

Well I know enough AI to answer this I guess.
I guess headscissors are a bit hard since they involve a bunch of factors like different positions, sexes, background,lighting... but with a big amount of data I believe you can do that. but a basic AI knowledge is not enough, since Computer Vision won't help in this kind of situation that easily, so I think building your own neural network is the safest/efficient option.

NOTE:
I'm building one to recognize female bodybuilders