The availability of image data online has created troves of visual data that we are only now starting to exploit. Recent advancements such as large-scale structure from motion algorithms has allowed the 3-D reconstruction of very large models (think: cities) solely from visual data downloaded off the Internet.

In this talk, we will introduce novel practical algorithms that leverage these reconstructions to reliably estimate the lighting conditions in image collections, while simultaneously estimating data-driven material representations of surfaces. This is done by exploiting a novel dataset of natural illumination, captured in conjunction with ground-level images of man-made landmarks.

This is joint work with Iain Matthews, from Disney Research. The presentation will be given in English.