3D imaging technology has recently made it possible to digitise a full-length human body with a high degree of accuracy and efficiency. The resulting 3D models provide a tremendous amount of information about the shape of the human body. In order to use these data to analyse the human body, one fundamentally important problem is to identify key anatomical points on the models. These points will help, among other things, to bring different human models into correspondence so that statistical analysis can be performed. Anthropometric landmarks are special points on the human body that specify well-defined and stable locations of anatomical parts. We address the problem of automatically identifying and locating anthropometric landmarks on 3D human models. Our method is based on statistical learning. Local surface properties and spatial relationships between landmarks are modeled as a pairwise Markov network. We will show that anthropometric landmarks can be located accurately through training and probabilistic inference.

Ce séminaire sera présenté à 11h00 dans la salle PLT-2704.