The Utility of 3D Landmarks for Arbitrary Pose Face Recognition

The authors investigate the utility of 3D facial landmark localisation in addressing the varying pose problem in 3D face recognition. They do not focus on the 3D landmark localisation problem itself, rather, they ask: Given the set of salient landmarks visible at some specific pose, what 3D face recognition performance can they expect, given that statistical training was performed at some other (canonical) pose? They perform standard Eigenface recognition for 3D data over a range of pose variations. The results indicate that localisation of facial landmarks is useful for pose-unconstrained face recognition although, as expected, recognition performance degrades as rotation away from the canonical training viewpoint increases.