Reference Points used for Arcface Alignment

Hi,

why do you use the following points for image alignment when doing preprocessing for training with arcface loss?

arcface_src = np.array([ [38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.7299, 92.2041] ], dtype=np.float32 )

It seems that first of all eyes and mouth points do not have the same y coordinates. Furthermore when doing mirroring along the y axis (which is a common data augmentation operation) the position of the landmarks changes slightly which could have a bad influence on training?

So where do these coordinates come from and why are they not designed such that the mirroring transformation does not change the landmark locations?

Kind regards,

Christian

[commanderka]

When the output image shall have size 224*224 the points used for alignment should be 2* np.array([ [38.2946, 51.6963], [73.5318, 51.5014], [56.0252, 71.7366], [41.5493, 92.3655], [70.7299, 92.2041] ], dtype=np.float32 ) Would be interesting to see how the choice of these alignment landmarks influences recognition performance, but I guess you need lots of computational resources to evaluate.

[nttstar]

It’s coming from another paper(git repo). Slight y-axis changes while doing left-right mirroring can bring some image-augmentation.