We propose a novel multiscale saliency detection algorithm for 3D meshes based on random walk framework. We construct a weighted undirected graph on an input 3D mesh model, by taking the vertices and edges in the mesh as the nodes and links of the graph. We compute a curvature value at each vertex using position and normal information, and assign a high weight to an edge connecting two vertices which have distinct curvature values each other. We perform random walk on the graph and find the stationary distribution of random walker, which is used as an initial saliency distribution. Moreover, in addition to local curvature characteristics, we also reflect global attributes of 3D geometry for saliency detection. We employ the saliency distributions at coarser scale meshes as restarting distributions of the random walker at finer scale meshes, based on random walk with restart framework. Experimental results show that the proposed algorithm detects the overall salient regions in 3D meshes as well as their local geometric details, faithfully.