Semantic and Generative Deep Learning of 3D Object Shapes
Supporting the below United Nations Sustainable Development Goals:支持以下聯合國可持續發展目標：支持以下联合国可持续发展目标：
ABSTRACT
Geometric deep learning is among the new trend of technologies that can model and learn from non-Euclidean data such as point clouds; it has recently been successfully used in data-driven solutions of 3D surface reconstruction, 3D semantic analysis, robotic manipulation, etc. In this talk, we introduce our recent progress along this line, including an analytic meshing technique that can exactly recover the surfaces from deep implicit networks, and sparse steerable convolution that enables efficient SE(3)-equivariant convolutions. We then introduce how these techniques can be used for robotic manipulation/grasping, autonomous driving, and other applications. We discuss the future potentials and challenges therein.
BIOGRAPHY
Kui JIA is a Professor with South China University of Technology. He received the B.E. degree from Northwestern Polytechnic University, the M.E. degree from National University of Singapore, and the Ph.D. degree from University of London. He was with Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences, Chinese University of Hong Kong, Advanced Digital Science Center, University of Illinois at Urbana-Champaign, and University of Macau. He has been regularly published in top-tier AI conferences and journals, such as TPAMI/CVPR/ICCV/NeurIPS/ICML. His recent research focuses on theoretical deep learning and geometric deep learning. His works won the ICDP Best Paper Award and were among the CVPR19 best paper finalists. He has been serving as an Associate Editor for IEEE. Trans. on Image Processing and as an Area Chair for NeurIPS.