Latent Gaussian Mixture Regression for Human Pose Estimation

Abstract

Discriminative approaches for human pose estimation model the functional mapping, or conditional distribution, between image features and 3D pose. Learning such multi-modal models in high dimensional spaces, however, is challenging with limited training data; often resulting in over-fitting and poor generalization. To address these issues latent variable models (LVMs) have been introduced. Shared LVMs attempt to learn a coherent, typically non-linear, latent space shared by image features and 3D poses, distribution of data in that latent space, and conditional distributions to and from this latent space to carry out inference. Discovering the shared manifold structure can, in itself, however, be challenging. In addition, shared LVMs models are most often non-parametric, requiring the model representation to be a function of the training set size. We present a parametric framework that addresses these shortcoming. In particular, we learn latent spaces, and distributions within them, for image features and 3D poses separately first, and then learn a multi-modal conditional density between these two low-dimensional spaces in the form of Gaussian Mixture Regression. Using our model we can address the issue of over-fitting and generalization, since the data is denser in the learned latent space, as well as avoid the necessity of learning a shared manifold for the data. We quantitatively evaluate and compare the performance of the proposed method to several state-of-the-art alternatives, and show that our method gives a competitive performance.