Evaluating error functions for robust active appearance models

Abstract

Active appearance models (AAMs) are generative parametric models commonly used to track faces in video sequences. A limitation of AAMs is they are not robust to occlusion. A recent extension reformulated the search as an iteratively re-weighted least-squares problem. In this paper we focus on the choice of error function for use in a robust AAM search. We evaluate eight error functions using two performance metrics: accuracy of occlusion detection and fitting robustness. We show for any reasonable error function the performance in terms of occlusion detection is the same. However, this does not mean that fitting performance is the same. We describe experiments for measuring fitting robustness for images containing real occlusion. The best approach assumes the residuals at each pixel are Gaussianally distributed, then estimates the parameters of the distribution from images that do not contain occlusion. In each iteration of the search, the error image is used to sample these distributions to obtain the pixel weights