A Comparative Study of Surface Representations Used in Statistical Human Models

This paper presents a quantitative and qualitative analysis of surface representations used in recent statistical models of human shape and pose. Our analysis and comparison framework is twofold. Firstly, we qualitatively examine generated shapes and poses by interpolating points in the shape and pose variation spaces. Secondly, we evaluate the performance of the statistical human models in the context of human shape and pose reconstruction from silhouette. The analysis demonstrates that body shape variation can be controlled with a lower dimensional model using a PCA basis in the Euclidean space. In addition, the Euclidean representation is shown to give more accurate shape estimates than other surface representations in the absence of pose variation. Furthermore, the analysis indicates that shape and pose parametrizations based on translation and rotation invariant representations are not robust for reconstruction from silhouette without pose initialization.