Multi-scale Spatial Offset-Attention Network for 3D Point Clouds Classification

Abstract

The method of extracting point cloud features has an important influence on the classification performance of 3D point clouds deep learning networks. 3D point clouds are disordered. However, the position relationship and spatial information between point pairs is stable, which can be used to extract feature information effectively. Similarly, the attention mechanism can capture feature information by learning the corresponding attention weight of various points. Therefore, a multi-scale spatial offset-attention network which fully utilizes spatial information of 3D point clouds is proposed. Particularly, the position relationship and spatial information between pairs are adopted to capture feature information effectively in the proposed module. The multi-scale network helps extract more comprehensive information. Moreover, for the purpose of reducing model complexity, we embed the residual structure into the input transform net and redesign the attention weight matrix in the spatial offset-attention module. Experiments on the proposed network which achieves excellent classification performance on challenge benchmark of 3D point clouds verify the ability of capturing point cloud features by fusing spatial information.