Deep Learning in the 3D World
Applying deep learning to 3D geometry is a challenging area of research because we typically represent 3D shapes as non-Euclidean data. This makes it difficult to apply convolutional neural networks (CNNs), because CNNs are designed to operate on Euclidean domains. Nevertheless, the deep learning community has devised many ingenious techniques to use deep learning for processing and generating 3D shapes. This deep dive session will cover the most exciting developments in the field of deep learning for 3D geometry. The following broad topics will be covered: 1) Using grid structures to create Euclidean domains for convolution. 2) Applying deep learning directly to unstructured 3D point clouds. 3) Generating 3D shapes using deep neural networks. Along the way, we will encounter popular 3D deep learning architectures such as PointNet and AtlasNet.
Prashant Raina is a Ph.D candidate at Concordia University, Montreal, completing his thesis titled “Fine Feature Reconstruction in Point Set Surfaces Using Deep Learning”. His research interests lie at the intersection of deep learning with 3D graphics, 3D geometry and computer vision. He is particularly interested in 3D point clouds as a geometry representation for both computer vision and rendering. He has previously given talks on his work on incorporating deep learning into 3D surface reconstruction from point clouds at the Autodesk Montreal Tech Talks in 2018, as well as the conference Graphics Interface 2018.