Reconstruction of animated models from images using constrained deformable surfaces

Abstract

In this paper we present a technique to generate animated 3D models from a limited set of images. We adopt a model-based approach to shape recovery to obtain a specific model parameterisation for animation. A generic animatable 3D mesh model is first manually posed to match the pose in the images. Shape information from multiple 2D image silhouettes is integrated by reconstructing the occupied volume of an object. The mesh is then deformed to match the reconstructed volume using a robust coarse to fine matching technique. We formulate an irregularly parameterised generic mesh as a shape constrained deformable surface to preserve the prior shape information and specific parameterisation in the model. Our interest lies in the generation of animated human models in a multi-camera studio. We assess the technique using artificial test cases and present results for a dancer in a studio.