To Frontalize or Not To Frontalize: A Study of Face Pre-Processing Techniques and Their Impact on Recognition

摘要

Face recognition performance has improved remarkably in the last decade. Muchof this success can be attributed to the development of deep learningtechniques such as convolutional neural networks (CNNs). While CNNs have pushedthe state-of-the-art forward, their training process requires a large amount ofclean and correctly labelled training data. If a CNN is intended to toleratefacial pose, then we face an important question: should this training data bediverse in its pose distribution, or should face images be normalized to asingle pose in a pre-processing step? To address this question, we evaluate anumber of popular facial landmarking and pose correction algorithms tounderstand their effect on facial recognition performance. Additionally, weintroduce a new, automatic, single-image frontalization scheme that exceeds theperformance of current algorithms. CNNs trained using sets of differentpre-processing methods are used to extract features from the Point and ShootChallenge (PaSC) and CMU Multi-PIE datasets. We assert that the subsequentverification and recognition performance serves to quantify the effectivenessof each pose correction scheme.