Improved Head-Shoulder Human Contour Estimation through Clusters of Learned Shape Models

In this paper we propose a clustering-based learning approach to improve an existing model for human head-shoulder contour estimation. The contour estimation is guided by a learned head-shoulder shape model, initialized automatically by a face detector. A dataset with labeled data is used to create the head-shoulder shape model and to quantitatively analyze the results. In the proposed approach, geometric features are firstly extracted from the learning dataset. Then, the number of shape models to be learned is obtained by an unsupervised clustering algorithm. In the segmentation stage, different graphs with an omega-like shape are built around the detected face, related to each learned shape model. A path with maximal cost, related to each graph, defines a initial estimative of the head-shoulder contour. The final estimation is given by the path with maximum average energy. Experimental results indicate that the proposed technique outperformed the original model, which is based on a single shape model, learned in a more simple way. In addition, it achieved comparable accuracy to other state-of-the-art models.