The CGV Colloquium will start online with a talk by Rana Hanocka (Tel Aviv University) on Friday, September 11 starting 16:00. The meeting is scheduled on Zoom, details are listed below.
Rana Hanocka is a 3rd-year PhD student in computer science at Tel Aviv University advised by Daniel Cohen-Or and Raja Giryes. At the intersection of Computer Graphics and Machine Learning, she is working on ways to use deep learning for manipulating, analyzing and understanding 3D shapes.
Title: Deep Learning on Single Shapes
Abstract:
One of the many difficulties in 3D deep learning is a lack of large, clean and labeled datasets. Acquiring and labeling large amounts of 3D data is not only cumbersome, but also requires using fundamental geometry processing pipelines that are not robust to geometry in the wild. Moreover, even high-quality 3D mesh models for similar shapes are extremely inconsistent with respect to triangulation and water-tightness, for example. On the other hand, training neural networks on single images has demonstrated surprisingly superb performance on a variety of different tasks. In this talk, I will present two recent works which propose training deep networks on a single shape. In Point2Mesh [SIGGRAPH 2020], we leverage the inductive bias of convolutional neural networks to learn a self-prior for surface reconstruction. We iteratively deform an initial mesh to “shrink-wrap” the input point cloud, resulting in a watertight mesh reconstruction. The weight-sharing property of CNNs models recurring and correlated structures within a single shape, and inherently removes noise and outliers. In Deep Geometric Texture Synthesis [SIGGRAPH 2020], we train a hierarchical GAN to learn to model the local geometric textures of a single shape. Our network displaces mesh vertices in any direction (i.e., in the normal and tangential direction), enabling synthesis of geometric textures, which cannot be expressed by a simple 2D displacement map.
*Zoom meeting details upon request.