Mesh saliency via ranking unsalient patches in a descriptor space.

This paper presents a novel mesh saliency detection approach based on manifold ranking in a descriptor space. Starting from the over-segmented patches of a mesh, we compute a descriptor vector for each patch based on Zernike coefficients, and the local distinctness of each patch by a center-surround operator. Patches with small or high local distinctness are named as background or foreground patches, respectively. Unlike existing mesh saliency methods which focus on local or global contrast, we estimate the saliency of patches based on their relevances to some of the most unsalient background patches, i.e. background patches with the smallest local distinctness, via manifold ranking. Compared with ranking with some of the most salient foreground patches as queries, this improves the robustness of our method and contributes to make our method insensitive to the queries estimated. The ranking is performed in the descriptor space of the patches by incorporating the manifold structure of the shape descriptors, which therefore is more applicable for mesh saliency since the salient regions of a mesh are often scattered in spatial domain. Finally, a Laplacian smoothing procedure is applied to spread the patch saliency to each vertex. Comparisons with the state-of-the-art methods on a wide range of models show the effectiveness and robustness of our approach.