ECCV 2020 Daily - Tuesday

2 Spotlight Presentation 8 Zhe Chen is a third-year PhD student at Kyoto University in Japan, under the supervision of Ko Nishino. His work is trying to solve a problem called inverse rendering. He speaks to us ahead of his spotlight presentation today. This work aims to estimate the reflectance and illumination from a single image of an object with known geometry. To do this, ideally, the reflectance and illumination would be easy to optimize and you would choose low-dimensional models for them. However, to ensure the estimation accuracy is unaffected by the expressive power of the models, they need to be high-dimensional, which is a contradiction. The solution is to have the best of both worlds by compacting the tractability of low-dimensional models and the expressiveness of high-dimensional models into a new invertible neural BRDF model composed of two parts. The invertible neural BRDF model is a conditional normalizing flow and it’s very expressive. In experiments, it is as expressive as the nonparametric bivariate BRDF model. Normalizing flow is an invertible neural network that transforms from a simple distribution to a complex distribution. By viewing the BRDF as a distribution function, it’s quite natural to use normalizing flow to model such a BRDF function. “The high-dimensional BRDF model is very expressive,” Zhe explains. “We would like it to be easy to optimize so we conditioned the normalizing flow model on a low-dimensional embedding space that captures the variety of real- world natural materials . Since this is a low-dimensional space, it’s really easy to optimize . That’s what we propose for the reflectance part.” Invertible Neural BRDF for Object Inverse Rendering DAILY T u s d a y