Face Refinement through a Gradient Descent Alignment Approach

Abstract

The accurate alignment of faces is essential to almost all automatic tasks involving face analysis. A common paradigm employed for this task is to exhaustively evaluate a face template/classifier across a discrete set of alignments (typically translation and scale). This strategy, provided the template/classifier has been trained appropriately, can give one a reliable but ``rough'' estimate of where the face is actually located. However, this estimate is often too poor to be of use in most face analysis applications (e.g. face recognition, audio-visual speech recognition, expression recognition, etc.). In this paper we present an approach that is able to refine this initial rough alignment using a gradient descent approach, so as to gain adequate alignment. Specifically, we propose an efficient algorithm which we refer to as the sequential algorithm, which is able to obtain a good balance between alignment accuracy and computational efficiency. Experiments are conducted on frontal and non-frontal faces.