Multi-Scale Attentional Network for Point Cloud Processing

Multi-scale architectures have found extensive applications in convolutional neural networks, enabling the extraction of more powerful feature representations. However, point clouds are discrete sets of 3D spatial data composed of individual points, making it challenging to perform multi-scale feature extraction using different convolution kernels. In this paper, we propose a novel neural network module that addresses this issue and allows for the integration of multi-scale features in point clouds. To compensate for the inherent lack of topological information in point cloud data, the module adopts a dynamic graph construction approach to gather more point features. This module efficiently obtains multi-scale features without incurring additional overhead. Moreover, it leverages an attention-like mechanism to effectively fuse these features, resulting in a more comprehensive and accurate representation, overcoming the drawbacks of traditional point-wise MLP followed by pooling operations that may lose valuable information. Experimental results demonstrate that our network outperforms existing methods in point cloud classification tasks on common benchmark dataset, achieving higher classification accuracy.