Learning disentangled representations of point clouds via alpha complexes for 3D shape classification

Abstract

Three-dimensional computer vision tasks have gained much attention in recent times, both in academic and industrial research. One of the key tasks of 3D computer vision is object classification. Various approaches based on the representations (e.g., point clouds, voxels, multi-view images and graphs) of the objects have been put forward for object classification. Recently, few works have used graph neural network for point cloud classification and have achieved promising results. In this thesis, we explore the use of a dual-stream graph neural network combining the alpha complexes constructed on the feature and non-feature regions of the point cloud object. The disentangled representation of the point cloud into feature and non-feature regions is achieved through a gradient structure analysis procedure and a Corner and Edge detection technique. Our experiments on ModelNet40 benchmark dataset indicate that the proposed graph-based method achieves higher or comparable accuracy to other state-of-the-art methods.