2016/10/16 by Sandipan Banerjee, Joel Brogan, Banerjee, Sandipan +13
Computer Science · #Biometric Identification and Security #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Face and Expression Recognition #Face recognition and analysis
paper · pdf · doi:10.48550/arxiv.1610.04823
openalex publication_date 2016/10/16 · openalex created_date 2022/10/03 · openalex updated_date 2026/07/28
Face recognition performance has improved remarkably in the last decade. Much\nof this success can be attributed to the development of deep learning\ntechniques such as convolutional neural networks (CNNs). While CNNs have pushed\nthe state-of-the-art forward, their training process requires a large amount of\nclean and correctly labelled training data. If a CNN is intended to tolerate\nfacial pose, then we face an important question: should this training data be\ndiverse in its pose distribution, or should face images be normalized to a\nsingle pose in a pre-processing step? To address this question, we evaluate a\nnumber of popular facial landmarking and pose correction algorithms to\nunderstand their effect on facial recognition performance. Additionally, we\nintroduce a new, automatic, single-image frontalization scheme that exceeds the\nperformance of current algorithms. CNNs trained using sets of different\npre-processing methods are used to extract features from the Point and Shoot\nChallenge (PaSC) and CMU Multi-PIE datasets. We assert that the subsequent\nverification and recognition performance serves to quantify the effectiveness\nof each pose correction scheme.\n