Recursive Shape and Pose Determination Using Deformable Model

Kevin Bailly, Maurice Milgram

Abstract: This paper adresses the problem of estimating shape and pose parameters of a 3D-non-rigid model from a single image. We introduce a new fast and robust method to minimize reprojection error between 3D model and feature image points. For this purpose, a recursive process is proposed. First, pose is roughly approximated using a rigid model. This permits to analytically determine shape parameters. Pose and shape are updated in an iterative way. Tests carried out on both synthetic and real data using the CANDIDE-3 parameterized mesh[1] are very promising.