|M.Sc Student||Yudin Eric|
|Subject||Improving Facial Expression Analysis via Intrinsic|
Normalization of Surfaces
|Department||Department of Computer Science||Supervisors||Professor Ron Kimmel|
|Professor Eran Yahav|
|Full Thesis text|
When comparing two or more sets of data it is prudent to first normalize them. This process typically involves subtracting the mean from each example and dividing by its standard deviation or variance. Such a practice brings all data examples to similar scale and range, and can also serve to reduce dimensionality. Further, if all data examples truly represent a single class, normalization can reduce their variability, without sacrificing variance between classes.
Here we employ a geometric framework to extend the concept of data normalization to the domain of functions that lie on manifolds. We pose normalization in this context as an embedding of all examples into manifolds nearly isometric to one another. Using novel geometric tools, we propose an implementation for the case of discretized functions on triangulated two-dimensional manifolds.
We apply the proposed Intrinsic Normalization technique to the task of automatic facial action unit detection. This problem has received much attention in the literature ever since Ekman and Friesen introduced the Facial Action Coding System (FACS) in an attempt to codify facial expressions via modular components called action units. By normalizing cross-subject examples to a common template face, we are able to improve the results of state-of-the-art action unit detection algorithms.