Improving 3D Facial Action Unit Detection with Intrinsic Normalization

Data normalization techniques are commonly used to reduce intra-class variance while preserving inter-class differences that aid in classification and regression tasks. Such practices bring all data examples to similar scale and range, and help to decrease the dimensionality of categorization problems. In this paper we describe a novel use of a geometric framework to extend the concept of data normalization to the domain of functions that lie on surfaces. In this context, normalization is posed as an embedding of all example functions into Riemannian manifolds nearly isometric to one another. Using geometric tools, we propose an implementation for the case of discretized functions on triangulated meshes. To demonstrate the proposed framework we apply it as a preprocessing step to the task of automatic facial action unit detection for depth maps from a 3D scanner. We show how this subject-independent representation produces more accurate action unit classification results. We also empirically demonstrate that utilizing this intrinsic normalization technique indeed improves action unit recognition performance.