Computer Vision News - February 2022

27 Adrian Dalca Have you thought about basing a challenge or a competition around it? That’s an interesting idea! I like that. We do have a general medical image registration challenge, Learn2Reg, organized by Alessa Hering, Lasse Hansen, Mattias Heinrich and me. What work would you like to be doing in the future? In medical imaging, a lot of the work starts from the problem, and we come up with a solution, but I like part of my work to be about evolving technology and how we can use it in our field. Students get really excited about that. Things like attention, or implicit representations. Then the idea is, can we use this technology in our work to start something new? That might sound a bit silly, like a hammer without a nail, but to me, it’s exciting! that’s faster and more accurate, but it’s enabled us to look at problems we couldn’t look at before. We have a follow-up paper aboutbuildingatlases-representativeimages of the brain - conditioned on attributes we care about like age. We couldn’t have done that without VoxelMorph. And that project with the hyperparameters, without VoxelMorph, we couldn’t build a network on top of it that estimated the effect of the hyperparameter. Have other people built on VoxelMorph? VoxelMorphisprettywell cited.Somepeople took it and added all kinds of properties to it. There are a few papers out there right now that use the ideas but replace the core architecture.. People applied it to domains I’d never heard of, but it was so interesting to get all these emails from people trying to use it! Adrian with Guha Balakrishnan (now Assistant Professor at Rice University) and Amy Zhao (now a researcher working on Oculus at Facebook Reality Labs)