Abstract: A visual similarity-based 3D model retrieval system is proposed in the paper,and this approach measures the similarity among 3D models by visual similarity.By the approach,we first obtain orthogonal projections of the 3D models which contain the depth information of the models,and the n we get similarity among 3D models through matching the shape similarity among orthogonal projections in terms of both Zernike descriptor and Reeb graph.The experimental results indicate that this visua-l based approach is robust against rotation transform,noise,model degeneracy etc.,and achieves satisfactory performance.