[RFC] Human Action Recognition

[rlleshi]

Hi guys, great project!

It was suggested on this Reddit post, that using a (sex) position classifier would be an interesting feature for this project.

I basically build one with a self-generated, small dataset. You can check it out here. I imagine that you guys must have huge datasets that could be used to properly train these deep learning models.

Let's discuss what can be expected out of these models!

[JediMaster25]

I would really like something like this, but the technology is not good enough to consider adding it to stash yet. The original repository could be turned into a CommunityScript with an optional argument to pass the location of the stash database and work with that.

[scruffynerf]

I just wrote the basics of a deepdanbooru tagger (using existing tf keras model) which does pictures.

The idea of video analysis is interesting. But with the price of GPUs still thru the roof (even a $60 Coral TPU is going for $300), it's the wrong time to buy one. So someone with one already could use this to write a recognizer for Stash. Won't be me.

[rlleshi]

FYI: These models could yield much better results with a better dataset & some more effort.

So the only hindrance would be the GPU(s).

[skier233]

Bumping on this. I feel it was dismissed too easily originally.

[DogmaDragon]

I would really like something like this, but the technology is not good enough to consider adding it to stash yet. The original repository could be turned into a CommunityScript with an optional argument to pass the location of the stash database and work with that.

I think so too, having a working plugin and building it up over time before making it part of the Stash is a great idea.

[skier233]

I agree a plugin is fine. I havent had a chance to get this working yet locally but the accuracy numbers are great and this is definitely the future even if the model isnt perfect yet.

[Maista6969]

Out of scope for Stash and can be closed