Residual Transformer Network for 3D Objects Classification

Abstract

Three-dimensional point cloud is an efficient and flexible representation of three-dimensional structures. Recently, neural network algorithm has shown superior performance in the tasks of three-dimensional point cloud classification and segmentation. However, the good results shown in these tasks were experimentally obtained using complete and aligned 3D point cloud data, whereas real-world 3D point cloud data is missing and unaligned. The key to the difficulty of learning non- aligned point cloud data is how to obtain invariance of geometric transformation. To solve this problem, we propose a new spatial transformation network named residual space transformation network to process point cloud data. Different from the existing spatial transformation networks, the network is inspired by traditional image and point cloud alignment algorithms and residual learning, and predicts the three-dimensional space transformation through a multi-step estimation method. The experiment shows that compared with the baseline, the network performs better in the classification and attitude prediction of missing point cloud data.