Multi-stream gaussian mixture model based facial feature localization

• Published in: 2008 IEEE 16th Signal Processing, Communication and Applications Conference pp. 1 - 4
• Main Authors: Kumatani, Kenichi, Ekenel, Hazim K., Hua Gao, Stiefelhagen, Rainer, Ercil, Aytul
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.04.2008
• Subjects: Accuracy; Face; Facial features; Feature extraction; Mouth; Nose; Shape
• ISBN: 9781424419982; 1424419980
• ISSN: 2165-0608
• DOI: 10.1109/SIU.2008.4632752

This paper presents a new facial feature localization system which estimates positions of eyes, nose and mouth corners simultaneously. In contrast to conventional systems, we use the multi-stream Gaussian mixture model (GMM) framework in order to represent structural and appearance information of facial features. We construct a GMM for the region of each facial feature, where the principal component analysis is used to extract each facial feature. We also build a GMM which represents the structural information of a face, relative positions of facial features. Those models are combined based on the multi-stream framework. It can reduce the computation time to search region of interest (ROI). We demonstrate the effectiveness of our algorithm through experiments on the BioID Face Database.