Face state detection & categorization?

[skywalker999u]

I noticed that training would look more decent (at least in preview) if it is trained with targeted pairs of data with similar face states (e.g. eyes opened/closed, mouth opened/closed, face direction, face partially obstructed/out of frame etc) Just a suggestion, if categorization after alignment can be added as a feature to facilitate a more targeted training, wouldn't it be more efficient as there is no need to kinda brute force train against 2 pairs of data with too much disparity?

[Clorr]

I was also thinking about that, just to have a report of face poses after extract. It would help people knowing if their training data contains same poses from src face and target face.

If you have some link to share with code, I would appreciate

[skywalker999u]

I have little to no experience in computer vision and python, so I don't know whether my search results would be useful. And sorry but I currently have no hard determination to contribute a lot to the project. I am just a user with programming background and a bit of knowledge about machine learning and opencv.

Face direction: https://www.learnopencv.com/head-pose-estimation-using-opencv-and-dlib/

Eye open/close detection: (the technique seems to be applicable for mouth too) https://www.pyimagesearch.com/2017/04/24/eye-blink-detection-opencv-python-dlib/

[IUsedToBeAPygmy]

Dfaker has written some code to compare Faceset A to Faceset B to see the (dis)similarity between images so you can get an idea of which faces of A do not have a representative face in B...

https://github.com/dfaker/df/blob/master/imageGapsScanner.py

[dfaker]

@IUsedToBeAPygmy Yep, it's in my main repo now, it just uses the MSE of the 2d face landmarks, nothing fancy but surprisingly effective, another reason to have the alignments saved rather than calculated on demand.

The warping in DF uses the same distance measure to warp an input face to have the proportions of the most similar opposing face.

[GuitarHero4]

In this article, under the section Training for Low-End GPU's, it says that it helps to split faces into frontal, three quarter and side views. But wouldn't that help for high end GPU's as well, as it should give a model that has a better fit?

I'd love to take a crack at adding a feature like this, but I am "not very well versed" in Python (read: never done Python). I have a feeling that the data is already there, since we already have the meshes and it should be a case of calculating the angle between the eyes or something like that. If someone could give me a pointer on where to get started I could try this. Though I think it might be semi-trivial for someone more experienced. :)