TSI-GCN: Translation and scaling invariant GCN for 3D point cloud analysis

Point cloud is a crucial data format for 3D vision, but its irregularity makes it challenging to comprehend the associated geometric information. Although some previous research has attempted to improve deep learning on point cloud and achieved promising results, they often overlook the robust shape descriptors of 3D targets, making them susceptible to translation and scaling transformations. This paper proposes a novel framework for point cloud analysis, to achieve feature extraction with translation and scaling invariance. It mainly includes local adaptive kernel, translation and scaling invariant convolution (TSIConv), and graph attention pooling. The key component is the design of TSIConv, which extracts the shape information with translation and scaling invariance. Then it performs convolution with local adaptive kernels to capture the features in various shape structures. Following the convolution layer, we add the graph attention pooling to coarsen point cloud, thus achieving multi-scale analysis and computational overhead reduction. The proposed framework, consisting of two networks, completes point cloud classification and part segmentation tasks in an end-to-end manner. The property analysis and experiments demonstrate that our model strictly guarantees the translation and scaling invariance, meanwhile achieving comparable performance to previous methods.