Abstract: This paper addressed the problem about the consistency of feature points,with different posed of 3D deformable shapes．It proposed a new algorithm to hierarchically detect feature points,based on the learning samples．Firstly,the algorithm detects the external feature points from input 3D shapes．Secondly,according to the local similarity of external points under different postures,semantic tags of external point are recognized by heat kernel signature and support vector machine．Finally,other feature points are hierarchically extracted by combining semantic tags and the geodesic distance of external points．In experiment,the proposed algorithm is proven to be very robust in detecting semantic-aware feature points on deformable shapes.

Key words: 3D model feature point,Hierarchical extraction,Semantic tags,Support vector machine,Heat kernel signature