Abstract: 3D shape segmentation is an important issue in shape analysis.This paper proposed a segmentation method based on multi-feature fusion to solve the consistency problem.A 3D model is first divided into multiple sub-patches by using over-segmentation.Then we use the geometric features extracted from each sub-patch as low-level features input for depth neural network model to generate high-level features.Finally,based on these high-level features,Gaussian mixture model is employed to get the clustering centers and graph-cut is adapted for the final segmentation.Experiments on PSB and COSEG datasets show that the proposed method outperforms the traditional geometric feature method,and can get good consistency results for the same kind of 3D shapes.