Feature-Based Pose Estimation

Abstract

In this chapter we review challenges and methodology for feature-based predictive three-dimensional human pose reconstruction, based on image and video data. We argue that reliable 3D human pose prediction can be achieved through an alliance between image descriptors that encode multiple levels of selectivity and invariance and models that are capable to represent multiple structured solutions. For monocular systems, key to reliability is the capacity to leverage prior knowledge in order to bias solutions not only to kinematically feasible sets, but also toward typical configurations that humans are likely to assume in everyday surroundings. In this context, we discuss several predictive methods including large-scale mixture of experts, supervised spectral latent variable models and structural support vector machines, asses the impact of the various choices of image descriptors, review open problems, and give pointers to datasets and code available online.