Computer Vision News - July 2020

learning techniques that we’re familiar with. During inference, this representation is also explicit because we have convex polytopes . We can just compute the convex polytopes so we can directly extract the surface. This is different from other implicit representations because for those representations we need some post- processing steps like Marching Cubes to actually get their surfaces. But for ours we can directly get them so it’s more efficient. It’s simpler. That’s very exciting to me.” There are many applications for this, including physics engines in games and autonomous driving cars . People working on physics simulation or collision detection like everything to be convex. With that property, the simulation will be much faster. This technique can be seamlessly incorporated in those pipelines to speed up the whole simulation. Thinking about next steps, Boyang says they are planning to put more research into an application- oriented framework. He is also keen to explore, now they have this good decomposition of shapes, can they have semantic parts on top of these simple convex primitives? We can’t let Boyang go without asking him what it is like to work with Geoffrey Hinton . Geoff is an author on this paper and Boyang tells us he brought him into the field of inverse “Andrea Tagliasacchi is a hands-on mentor teaching me how to do good research!” Oral Presentation 24 Best of CVPR 2020