Mixture of Ridge Regressors for Human Pose Estimation

Mixture models have been popular to solve computer vision problems such as optical flow, object recognition and human pose estimation. In particular, mixtures of classifiers are state-of-the-art approaches to estimate human pose from images. These discriminative approaches learn a functional mapping, or conditional distributions, between image features and 3D poses. However, existing algorithms to learn the parameters of the mixture model are prone to local minima, typically resulting in over-fitting and poor generalization. To partially address this problem, this paper proposes a new mixture model called mixture of ridge regressors (MoRRs). To learn the parameters of a MoRR, we propose a loss function that incorporates pair-wise smoothness constraints. These constrains along with a deterministic annealing optimization strategy improves the parameter estimation process relative to existing approaches. Experimental results on body and head pose estimation from images illustrate the benefits of the proposed approach with respect to state-of-the-art methods on standard databases.