Reconstructing Human Shape and Motion from Multi-View Video
Edilson de Aguiar, Christian Theobalt, Marcus A. Magnor, Hans‐Peter Seidel · 2005
In model-based free-viewpoint video, a detailed representation of the time-varying geometry of a real-word scene is used to generate renditions of it from novel viewpoints. In this paper, we present a method for reconstructing such a dynamic geometry model of a human actor from multi-view video. In a two-step procedure, first the spatio-temporally consistent shape and poses of a generic human body model are estimated by means of a silhouette-based analysis-by-synthesis method. In a second step, subtle details in surface geometry that are specific to each particular time step are recovered by enforcing a color-consistency criterion. By this means, we generate a realistic representation of the time-varying geometry of a moving person that also reproduces these dynamic surface variations.