Point Cloud Classification Based on Separable Transformer

Transformer tends to take advantage of capturing remote dependencies to extract relational interactions at remote points of the point cloud, ignoring important local structural details, and achieves high performance by significantly increasing the computational burden. To alleviate this problem, we propose a separable Transformer point cloud classification method, named Sep-point, based on the idea of separable visual Transformer. The proposed Sep-point facilitates sequential local-global relational interactions within and between groups of point clouds through depth-separable self-attention. New location token embedding and group self-attention methods are used to compute inter-group attentional relationships with negligible computational cost and to establish telematic interactions across multiple regions, respectively. In this way, the local-global features are extracted while the computational burden is significantly reduced. Experimental results show that the proposed Sep-point improves the classification accuracy by 0.2% on the ModelNet40 dataset over the existing PCT (Point Cloud Transformer) and by 6.3% on the real ScanObjectNN dataset, respectively. Moreover, the number of network parameters and FLOPS metrics are reduced by 0.72M and 0.18G, respectively. These experimental results clearly demonstrate the promising effectiveness of our proposed method.