Graph based semi-supervised human pose estimation: When the output space comes to help

In this letter, we introduce a semi-supervised manifold regularization framework for human pose estimation. We utilize the unlabeled data to compensate for the complexities in the input space and model the underlying manifold by a nearest neighbor graph. We argue that the optimal graph is a subgraph of the k nearest neighbors (k-NN) graph. Then, we estimate distances in the output space to approximate this subgraph. In addition, we use the underlying manifold of the points in the output space to introduce a novel regularization term which captures the correlation among the output dimensions. The modified graph and the proposed regularization term are utilized for a smooth regression over both the learned input and output manifolds. Experimental results on various human activities demonstrate the superiority of the proposed algorithm compared to the current state of the art methods.

► Graph-based regression models are employed for the human pose estimation problem. ► We argue that the optimal graph is a subgraph of a k-NN graph. ► Distance of the points in estimated output space is used to find the optimal graph. ► In addition, output space manifold is used to introduce a new regularization term. ► Experimental results show significant improvement with respect to the existing methods.