Anisotropic diffusion map based spectral embedding for 3D CAD model retrieval

In the product life cycle, design reuse can save cost and improve existing products conveniently in most new product development. To retrieve similar models from big database, most search algorithms convert CAD model into a shape descriptor and compute the similarity two models according to a descriptor metric. This paper proposes a new 3D shape matching approach by matching the coordinates directly. It is based on diffusion maps which integrate the rand walk and graph spectral analysis to extract shape features embedded in low dimensional spaces and then they are used to form coordinations for non-linear alignment of different models. These coordinates could capture multi-scale properties of the 3D geometric features and has shown good robustness to noise. The results also have shown better performance compared to the celebrated Eigenmap approach in the 3D model retrieval.