MFRIN: Rotation-Invariant Network with Multi-feature Fusion of Point Cloud

Abstract

Point cloud carries rich geometric information and has unique advantages in computer vision field. Existing methods can effectively identify objects in a fixed perspective, but in practical, the object direction is unknown, which greatly affects the accuracy of the network. In this paper, we propose a pre-network based on ellipsoid fitting algorithm to extract rotation invariant features of point cloud. We design a two-branch network to mine features at different levels. In the first branch, we feed the rotation invariant feature to PointNet-based backbone to learn global feature; in the side branch, we use knn to transform the point cloud into a graph structure and apply attention model to extract more discriminative local features. We show that MFRIN can extract rotation-invariant representations for point cloud without data augmentation, and achieves best performance than state-of-the-art methods on rotating point clouds.