Recovering Illumination from a Single Image

ECE Seminar: Recovering Illumination from a Single Image


Starts at: December 7, 2017 4:30 PM

Ends at: 6:00 PM

Location: SH 125

Speaker: Dr. Kalyan Sunkavalli

Affiliation: Senior Research Scientist

Link to Abstract

Link to Video (1)

Details:We see the world because it is lit by illumination from all around us. Estimating this illumination is critical for a number of vision and graphics tasks like scene understanding, image editing, and augmented reality. However, inferring illumination from a single image is an extremely challenging problem. We often do not actually see lighting in images; we only observe it indirectly via its interactions with the unknown shapes and materials in the scene. In this talk, I will present two attempts at tackling this problem for outdoor and indoor scenes. In both cases, we have developed a way to create training data from a large-scale LDR panorama dataset. Using this data in conjunction with a physics-motivated training process allows us to train deep neural networks to predict HDR illumination from a single LDR image. We are able to produce state-of-the-art illumination estimates, and some of this work is now part of the 3D compositing tool, Adobe Dimension CC.