ICCV Daily 2021 - Thursday

essence of computer vision, but the techniques can be applied elsewhere. Everything this paper does is about the geometry and the optics of image formation, but then it is signal processing and linear algebra from that point onwards. Thinking about next steps, Jon tells us there are still limitations to NeRF. The models are slow and making them fast is an ongoing problem to solve. Also, NeRF bakes in all the lighting effects, so you can move the camera, but you cannot change the light or the textures. In the future, Jon would like NeRF to be a general-purpose inverse rendering tool that can do relighting, ray tracing, material editing, and all these things you would expect from sophisticated graphics engines. “ This I feel is step one of at least a 10-step process of modernising NeRF, ” he adds. “ NeRF is simple in many ways. It uses primitive technology, ray tracing, and fully connected neural networks. Everything we’ve been doing is from the ‘ 80s and we’re slowly bringing NeRF into the ‘ 90s and the ‘ 00s by adding more sophisticated things that are well understood in graphics but haven’t ported over to computer vision and deep learning. We’re still on aliasing, which is a simple concept, but making these things neural is challenging. ” Jon tells us his tastes have always leaned towards older papers from the ‘ 80s and ‘ 90s. “ Things were simpler back then, ” he adds. “ People were focused on pixels and low-level things, which allowed for simple mathematical treatments and elegant algorithms. I’m excited to use my knowledge and enthusiasm for old computer vision and image processing papers and apply them in a modern context. ” To learn more about Jon’s work [Paper ID 4397], you can view his pre- recorded oral presentation and PDF poster now and join his live Q&A session today [5B] at 09:00-10:00 EDT. There is also a project page. 7 DAILY ICCV Thursday Jon tells us he has seen evidence that NeRF is a kind of general-purpose tool for many things outside of computer vision, from computer graphics to robotics. The system is designed to take in images and to produce models from that, which is the Jon Barron

RkJQdWJsaXNoZXIy NTc3NzU=