Recent work on outdoor illumination estimation made use of existing analytical sky models. However, these sky models can only represent a small subset of the vast diversity of the outdoor lighting appearance. In this presentation, we will discuss our novel data-driven and learned sky model and describe a method to perform illumination estimation from a single image, solving both problems of lighting modelling and estimation end-to-end. Finally, we will demonstrate the versatility and expressivity of our learned sky model and show that it can be used to recover plausible illumination from a single limited field of view image, leading to visually pleasant virtual object insertions.

Biography

Yannick Hold-Geoffroy is a Research Engineer in the Creative Intelligence Lab with Adobe Research in San Jose. He received his Ph.D. degree in Electrical Engineering at Université Laval, Canada, in 2018. His research interests are in the understanding of natural images through machine learning and lighting modelling.

The seminar will be presented at 11:00 a.m. in room PLT-3370.