Modeling of pose effects in oriented filter responses for head pose estimation

Abstract

We propose an approach for view angle invariant recognition of 3D objects, based on modeling the variations of local feature values as function of view angle. In recognition stage we can compute the probabilities for any pixel that there is certain feature in a given pose angle. Any maximum likelihood or posterior based estimation methods can then be applied to infer the objects and their view parameters. We demonstrate the method with piecewise linear model for the pose effects, to recognize the location and pose of a head from the two eyes. © Springer-Verlag 2003.