Inverse Rendering

One of the main goals in the image processing research, from the early days until now, is the ability to describe the scene in terms of “intrinsic” characteristics like depth, shape, surface orientation (normals), incident light, and reflectance at each visible surface point. Each of these intrinsic characteristics provides us with valuable cues for scene understanding. Accurate intrinsic scene characterization is essential in various tasks including: photo forensics, segmentation, object classification and recognition, scene relighting and automatic white balance, shadow removal, augmented reality, object recoloring, and etc.
Inverse Rendering refers to estimation of intrinsic scene characteristics given a single photo or a set of photos of the scene. While predicting these characteristics from 2D projects is highly underconstrained, recent advances in this topic have made a big step in solving this problem. Inverse rendering methods have already made their way into end-user applications.
Our latest achievement in this field is our novel multi-illuminant dataset published in ICCV 2015.