A Novel Shape Registration Framework and Its Application to 3D Face Recognition in the Presence of Expressions

Abstract

This paper introduces a new global-to-local shape registration technique and shows its potential in solving the problem of 3D face recognition in the presence of expressions. The proposed registration technique is a two-step technique that operates in an implicit higher dimensional space where the powerful distance transform is used as the embedding function. First, a new dissimilarity measure is introduced to recover the transformation that globally aligns the two input shapes. This new measure can deal efficiently with rigid, similarity and affine motions. Second, the local coordinate transformation between the two globally aligned shapes is explicitly estimated by minimizing a new energy functional consisting of three terms. The first term is a discrepancy measure between the two shape representations. The second term penalizes the deviation of the distance map representation of the globally warped source shape from a signed distance function, while the local displacement field is being updated. The last term is a regularization term that enforces the smoothness of the recovered deformations. This leads to a set of coupled equations that are simultaneously minimized through a gradient descent scheme. The overall potential of the proposed framework is demonstrated through various 2D/3D experimental results. As an application, we address the 3D face recognition problem in presence of facial expressions.