See #66 - this version is based on the master rather than pytorch branch:
Following the change from the previous face detector that would automatically axis-align the face to one that doesn't, the eye patches have become inconsistent as the roll of the head changes significantly from zero.
This branch corrects that by affine transforming the face image based on eye landmarks to better extract the eye patches.
It's done at minimal cost; the standalone and the ros nodes continue to work as expected
See #66 - this version is based on the master rather than pytorch branch:
Following the change from the previous face detector that would automatically axis-align the face to one that doesn't, the eye patches have become inconsistent as the roll of the head changes significantly from zero.
This branch corrects that by affine transforming the face image based on eye landmarks to better extract the eye patches.
It's done at minimal cost; the standalone and the ros nodes continue to work as expected