3D human pose estimation from single images using dual-wing harmonium model

Abstract

In this paper, we propose a metric learning scheme for 3D human pose estimation from single images based on dual-wing harmonium model (DWH). Specifically, we focus on a big challenge of this problem which is different 3D poses may correspond to similar inputs features. Based on the dual-wing harmonium model, our method provides a principled way to embed two-modal data (input features and pose features) into a single latent space such that, in that space the data points which have similar poses are near each other. For estimating the pose feature for a new data point, we first determine the latent feature of that data point, then, we simply assign the mean of the pose features corresponding to the nearest neighbours of that data point in the new space. The experimental results demonstrate that the performance of the proposed method is significantly better than the state of the art 3D human pose estimation methods.