Corresponding dynamic appearances

摘要

Modelling the appearance of 3D objects undergoing large pose variation relies on recovering correspondence of both shape and texture across views. The problem is hard because changes in pose not only introduce self-occlusions hence inconsistent 2D features between views, but also cause non-linear variations in both the shape and texture of object appearance. In this paper, we present an approach for establishing structured sparse correspondence between face images across views using non-linear shape models. We extend the non-linear shape models to dynamic appearance models of both shape and texture across views. For non-linear model transformation, we adopt Kernel PCA. For bootstrapping appearance alignment at different views, we introduce a generic-view shape template. We show that Kernel PCA constrained the dynamic appearance model and eases model fitting to novel images.