Clustered Pose and Nonlinear Appearance Models for Human Pose Estimation

Human pose estimation is the task of estimating the ‘pose’ or configuration of a person’s body parts e.g. labeling the position and orientation of the head, torso, arms and legs in an image. In this paper we propose an extension of the pictorial structure model (PSM) approach [2]. Our method incorporates richer models of appearance and prior over pose without introducing unacceptable computational expense. We build on the idea of a ‘mixture of trees’ model [3]. The space of human poses is partitioned into a set of clusters such that the prior over plausible poses can be modeled with greater fidelity. Within each pose cluster we use pose-specific appearance terms which implicitly capture the dependence of a part’s appearance on pose and the correlation between the appearance of parts. To cope with the large variation in part appearance due to factors such as clothing or varying anatomy we use state-of-the-art nonlinear SVM classifiers to model the appearance terms. This would typically be prohibitive in terms of computational expense, however we show that by adopting a cascaded reduced set machine formulation [6] we can exploit such strong classifiers efficiently. Current methods have been limited by the lack of available training data – to overcome this we introduce a new annotated dataset of 2,000 diverse and challenging consumer images which will be made publicly available (Figure 1). Our results show that the use of stronger appearance terms and prior model in the proposed approach results in a greater than 50% relative improvement in pose estimation accuracy on this dataset compared to a state-of-the-art method [4].