WEIWEI GUO, IOANNIS PATRAS: LOKD REGRESSION FOR HUMAN POSE ESTIMATION 1 Learning Output-Kernel-Dependent Regression for Human Pose Estimation

This paper presents a novel method for continuous inter-dependent outputs prediction that predicts each of the multiple output variables based not only on the input but also on the rest of the outputs. We do so by using for each output kernel regression functions that are a convex combination of two kernels: the first kernel defined over the input and the second over the remaining outputs. We propose a scheme in which the relative importance of these two heterogeneous information sources is learned jointly with the parameters of the regression scheme. The inference for a new observation is obtained by fix-point iteration which is based on alternatively evaluating the learned outputdependent functions until convergence. We experimentally validate our proposed method on 3D human pose estimation problem using the HumanEva-I dataset.