IMPROVED FACE FINDING IN VISUALLY CHALLENGING ENVIRONMENTS (WedAmOR5)

Author(s) : Jintao Jiang (House Ear Institute, United States of America) Gerasimos Potamianos (IBM T.J. Watson Research Center, United States of America) Giridharan Iyengar (IBM T.J. Watson Research Center, United States of America)

Abstract : Finding faces in visually challenging environments is very crucial to many applications such as audio-visual ASR, video indexing, user interface, and video surveillance. In this study we investigate several algorithms to improve face detection performance in visually challenging environments under the IBM face detection framework. These algorithms are a trainable skintone pre-screening, Hamming windowing of face images, DCT component selection, and AdaBoost. When these algorithms are combined, a 15% improvement in face detection for one of the challenging datasets is observed. The results show that a robust face detection system can be achieved with consideration of various factors, and the combination of DCT component selection and AdaBoost algorithms comprise a feature selection and multi-template system.

Menu