3D Shape Segmentation via Shape Fully Convolutional Networks

We propose a novel fully convolutional networks architecture for shapes, denoted as Shape Fully Convolutional Networks (SFCN). Similar to convolution and pooling operation on image, the 3D shape is represented as a graph structure in the SFCN architecture, based on which we first propose and implement shape convolution and pooling operation. Meanwhile, to build our SFCN architecture in the original image segmentation FCN architecture, we also design and implement the generating operation with bridging function. This ensures that the convolution and pooling operation we designed can be successfully applied in the original FCN architecture. In this paper,we also present a new shape segmentation based on SFCN. In contrast to existing state-of-the-art shape segmentation methods that require the same types of shapes as input, we allow the more general and challenging input such as mixed datasets of different types of shapes. In our approach, SFCNs are first trained end-to-end, triangles-to-triangles by three low-level geometric features. Then, based on the trained SFCNs, we can complete the shape segmentation task with high quality. Finally, The feature voting-based multilabel graph cuts is adopted to optimize the segmentation results obtained by SFCN prediction. The experiment results show that our method can effectively learn and predict mixed shape datasets of either similar or different characters, and achieve excellent segmentation results.