Semi-Automatic Prediction of Landmarks on Human Models in Varying Poses

Abstract

We present an algorithm to predict landmarks on 3D human scans in varying poses. Our method is based on learning bending-invariant landmark properties. We also learn the spatial relationships between pairs of landmarks using canonical forms. The information is modeled by a Markov network, where each node of the network corresponds to a landmark position and where each edge of the network represents the spatial relationship between a pair of landmarks. We perform probabilistic inference over the Markov network to predict the landmark locations on human body scans in varying poses. We evaluated the algorithm on 200 models with different shapes and poses. The results show that most landmarks are predicted well.