represent array of 68 2D points as histogram to filter faces?

[iperov]

Is there way to represent 68 points alignments as histogram like

will it help to cut unnecessary faces to make faceset more regular and better for nn?

[Clorr]

I think this is what was the 1st intention in deepfakes/faceswap#123 . I agree it would be good to have this to be able to compare 2 facesets before starting training

[Clorr]

There are various examples around. I don't know if you have found some interesting ones. Maybe this one can help: https://github.com/mpatacchiola/deepgaze/blob/master/examples/ex_dlib_pnp_head_pose_estimation_video.py (One guy tried here to output the values as angles: mpatacchiola/deepgaze#40)

(Note that we can use the landmarks we already have instead of using their lib)

[iperov]

solved