Matching 3D Faces with Partial Data

We present a novel approach to matching 3D faces with expressions, deformations and outliers. The matching is performed through an accurate and robust algorithm for registering face meshes. The registration algorithm incorporates prior anthropometric knowledge through the use of suitable landmarks and regions of the face to be used with the Iterative Closest Point (ICP) registration algorithm. The localization of landmarks and regions is achieved through the fitting of a 3D Point Distribution Model (PDM) and is independent of texture, pose and orientation information. We show that the use of expression-invariant facial regions for registration and similarity estimation outperforms the use of the entire face region. Evaluation is performed on the challenging GavabDB database and we achieve 93.7% rank-1 recognition with an overall retrieval accuracy of 91.1%. The matching of 3D face meshes requires accurate comparison of surface properties from different meshes. This task can be impaired when meshes have deformations (due to expressions or medical conditions) or outliers (due to the acquisition process). A common approach to match 3D meshes is through the Iterative Closest Point (ICP) algorithm [3] for rigid registration. As ICP is based on the closest point associations from one mesh to the other, in the presence of deformations and outliers its performance degrades due to global minima. Moreover, ICP requires roughly aligned meshes in order to converge in terms of the mean square error (MSE). Many variants to the original ICP algorithm have been proposed to improve speed and convergence [2, 22], but without removing the above mentioned limitations. Other global registration methods exist [8, 9, 11], some of which use the ICP, but are also inappropriate in the presence of deformations and outliers. The localization of specific anthropometric landmarks and regions on face meshes often plays an important part in these applications. Landmarks can aid the ICP algorithm in achieving rough alignment of meshes, and by themselves provide valuable semantic information. In biometric applications, landmarks are often instrumental in the generation of signatures for faces [19] and isolation of expression invariant regions for matching [4]. However, the dependence on prior knowledge of feature map thresholds, orientation and pose is evident in most existing methods for landmark localization on meshes [4, 5, 14]. The segmentation of faces is also important prior to analysis when the mesh includes other