Computer Vision News - July 2020

Boyang Deng is an AI resident in the Google Brain Toronto team. His paper is about a differentiable shape representation based on convex polytopes. He spoke to us ahead of his oral presentation. This work presents a novel way to deal with large-scale shape collections . It’s a new bottom-up perspective that assembles shapes by putting a collection of primitives together. When shape representation has the property of convexity, it’s useful in graphics applications and real-world applications like collision detection and physics simulation . Convexity is a favourable property because it speeds up the whole process. When the representation is differentiable, it can be learned from a Oral Presentation 22 large-scale shape collection without very much human intervention. “We use an auto-encoding framework to learn a model that can decompose any shape into a collection of convexes,” Boyang explains. “To represent convex shapes, we use the idea that any convex can be represented by the intersection of a lot of half-spaces or half-planes in 3D. We also make use of the differentiability of indicator functions represented as sigmoid functions. Connecting the dots, we use those facts to build up a unified framework that can learn convex composition from data.” Boyang says it was Professor Alan Yuille from Johns Hopkins University who first got him thinking about this subject. He did a research internship with him back in 2017. “He was really nice to me and taught me many concepts about how to do research,” he says. “He was the first to plant the idea that by representing any visual objects we should have this bottom- up framework. We should have a lot of low-level concepts and a combinative or compositional model to organize them.” CvxNet: Learnable Convex Decomposition “[Geoff Hinton] wants to push this concept of vision as inverse graphics and I work closely with him on that. He’ll give me a high-level blueprint about how things should be in four-or five-years’ time!” Best of CVPR 2020