Abstract: Deep representation of 3D model is the key and prerequisite for 3D object recognition and 3D model semantic segmentation, providing a wide range of applications ranging from robotics, automatic driving, virtual reality, to remote sensing and other fields. However, convolutional architectures require highly regular input data formats and most researchers typically transform point cloud data to regular 3D voxel grids or sets of images before feeding them to a deep net architecture. The process is complex and the 3D geometric structure information will be lost. In this paper, we make full use of the existing deep network which can deal with point cloud data directly, and propose a new algorithm that uses double symmetry function and space transformation network to obtain more robust and discriminating features. The local topology information is also incorporated into the final features. Experiments show that the proposed method solves the problem of lacking local information in an end-to-end way and achieves ideal results in the task of 3D object recognition and 3D scene semantic segmentation. Meanwhile, the method can save 20% training time compared to PointNet++with the same precision.

Key words: point cloud, deep learning, raw data, 3D object recognition, 3D model segmentation