CVPR Daily - Wednesday

DAILY Wednesday Paul Henderson 7 modelling software, but instead, they could say ‘create me a car’ and the computer would generate it for them . Paul met his co-author, Vagia Tsiminaki , at ETH Zürich when he spent six months there during his PhD . “I knew that she was interested in reconstruction and something similar to differentiable rendering,” Paul tells us. “I come more from the machine learning probabilistic modelling side, so it was natural to talk to her and get her insights on the classical 3D reconstruction side.” The other co-author is his PI at IST Austria, Christoph Lampert , who he thanks for all his sage advice. To take this paper forward, Paul thinks the natural direction is to relax some of the assumptions it makes. One example he would like to relax is the assumption of an approximate camera calibration . He would also like to move from generating single objects to generating full coherent scenes , such as a busy office environment, without any 3D supervisio n. What does Paul find most exciting about this work? “I like this principle of embedding prior knowledge about how the world is built into the deep learning models,” he tells us. “In this case, it’s knowledge about how the 3D structure of the world gives rise to 2D images. Neural networks are just big black boxes normally, and I like this concept of giving them some structure instead. Hopefully, it will make them learn in a more human-like manner .” To find out more about Paul’s work, you are invited to watch his pre-recorded oral video at 14:30 (Wednesday) or 02:30 (Thursday) and visit one of the Oral 2.3B Q&A sessions at 14:00- 16:00 (Wednesday) or 02:00- 04:00 (Thursday).