Unsupervised Representation Learning of Point Cloud via Local Discriminability Enhancing

Capturing and annotating 3D point cloud data are very time-consuming and expensive. The unsupervised method is an effective solution to learn point cloud representations without data annotation. Most existing unsupervised methods take into account the importance of both local structure features and global features and combine them to learn object representations. However, they rarely consider that different local structures contribute differently for object representation. In this paper, we argue that the discriminative local structures are significant for object representation. Therefore, we propose a discriminability enhancement scheme to mine discriminative local structure features and further enhance their discriminability. Our unsupervised method can learn powerful representations of point clouds by fusing discriminative local structure features and global features. Experimental results show that our method can achieve superior performance on downstream classification tasks.