HIGCNN: Hierarchical Interleaved Group Convolutional Neural Networks for Point Clouds Analysis

Abstract

Although previous works for point clouds analysis have achieved remarkable performance, it is difficult for them to achieve a good trade-off between accuracy and complexity. In this paper, we present an efficient and lightweight neural network for point clouds analysis, named HIGCNN, which can achieve better performance but lower complexity compared to existing methods. The key component in our approach is the hierarchical interleaved group convolution (HIGConv) operation. We first present a neighborhood attention convolution (NAC) operation to fully mine fine-grained local geometric features inside each local area. With the proposed NAC, we further design a HIGConv to encode both fine-grained local geometric features and discriminative nonlocal point-wise features with fewer parameters and lower computational costs. To further capture fine-grained contextual features, we propose a multi-scale relation (MSR) module to fully explore the relationship among different scale areas. Extensive experiments show that our HIGCNN surpasses state-of-the-art approaches for classification and semantic segmentation on four benchmarks ModelNet40, S3DIS, vKITTI and SemanticKITTI in terms of accuracy and complexity.