2019/08/24 by Sameera Ramasinghe, Salman A. Khan, Ramasinghe, Sameera +5
Computer Science · Engineering · #3D Shape Modeling and Analysis #FOS: Computer and information sciences #Image Processing and 3D Reconstruction #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Medical Image Segmentation Techniques
paper · pdf · doi:10.48550/arxiv.1908.10209
openalex publication_date 2019/08/24 · openalex created_date 2022/09/27 · openalex updated_date 2026/07/28
Existing networks directly learn feature representations on 3D point clouds\nfor shape analysis. We argue that 3D point clouds are highly redundant and hold\nirregular (permutation-invariant) structure, which makes it difficult to\nachieve inter-class discrimination efficiently. In this paper, we propose a\ntwo-faceted solution to this problem that is seamlessly integrated in a single\n`Blended Convolution and Synthesis' layer. This fully differentiable layer\nperforms two critical tasks in succession. In the first step, it projects the\ninput 3D point clouds into a latent 3D space to synthesize a highly compact and\nmore inter-class discriminative point cloud representation. Since, 3D point\nclouds do not follow a Euclidean topology, standard 2/3D Convolutional Neural\nNetworks offer limited representation capability. Therefore, in the second\nstep, it uses a novel 3D convolution operator functioning inside the unit ball\n( mathbbB3) to extract useful volumetric features. We extensively derive\nformulae to achieve both translation and rotation of our novel convolution\nkernels. Finally, using the proposed techniques we present an extremely\nlight-weight, end-to-end architecture that achieves compelling results on 3D\nshape recognition and retrieval.\n