Face alignment is an essential task in computer vision and is of vital importance in face recognition. In this work, face registration makes a leap by transitioning from a simple eye-based alignment method to a dense, landmark-based alignment. Mohammadzade et al. show that face recognition achieves a boost in performance when a non-linear method, constrained by landmark points, is employed for face alignment. Compared to regular alignment, usually based on two fiducial points located on the eyes, their method achieves pixel-wise alignment by means of a piecewise, locally linear landmark-based transformation. This novel alignment technique is tightly coupled with linear discriminant analysis classifiers working on face patches, encoding both geometry and appearance information. The proposed method shows remarkable performance not only in a controlled environment but also on imagery captured “in the wild.”
You must log in
to add comments.