188宝金博页面版

Hierarchical Shape Modeling for Automatic FaceLocalizationCe Liu1, Heung-Yeung Shum1, and Changshui Zhang21Visual Computing Group, Microsoft Research Asia, Beijing 100080, China2Department of Automation, Tsinghua University, Beijing 100084, Chinalce@msrchina. research. microsoft. comAbstract. Many approaches have been proposed to locate faces in animage. There are, however, two problems in previous facial shape modelsusing feature points. First, the dimension of the solution space is toobig since a large number of key points are needed to model a face. Sec-ond, the local features associated with the key points are assumed to beindependent. Therefore, previous approaches require good initialization(which is often done manually), and may generate inaccurate localiza-tion. To automatically locate faces, we propose a novel hierarchical shapemodel (HSM) or multi-resolution shape models corresponding to a Gaus-sian pyramid of the face image. The coarsest shape model can be quicklylocated in the lowest resolution image. The located coarse model is thenused to guide the search for a finer face model in the higher resolutionimage. Moreover, we devise a Global and Local (GL) distribution tolearn the likelihood of the joint distribution of facial features. A novelhierarchical data-driven Markov chain Monte Carlo (HDDMCMC) ap-proach is proposed to achieve the global optimum of face localization.Experimental results demonstrate that our algorithm produces accuratelocalization results quickly, bypassing the need for good initialization.1IntroductionFace detection and face localization have been challenging problems in computervision and machine perception. Face detection, for example, explores possiblelocations of faces from an input image, and face localization accurately locatesthe facial shape and parts, often from an initialized model. Appearance modelshave been successfully used for face detection, where typically a square regionwith an elliptic mask is used to represent a face image. Based on a large amountof positive (face) and negative (non-face) samples, machine learning techniquessuch as PCA [13], neural networks [9,11], support vector machines [6], wavelets[10] and decision trees [14], are always used to learn the separating manifold offaces and non-faces. By verifying patterns in a shifting window, the position ofa face can be derived.However, an appearance model alone is not flexible enough to model shapedeformations and pose or orientation variations. Shape models, in particulardeformable shape models such as deformable template matching [15] and graphA. Heyden et al. (Eds.): ECCV 2002, LNCS 2351, pp. 687–703, 2002.c Springer-Verlag Berlin Heidelberg 2002