Higher Order Priors for Joint Intrinsic Image, Objects, and Attributes Estimation

摘要: Many methods have been proposed to solve the problems of recovering intrinsic scene properties such as shape, reflectance and illumination from a single image, object class segmentation separately. While these two are mutually informative, in past not many papers addressed this topic. In work we explore joint estimation recovered an together with objects attributes present scene. way, our unified framework is able capture correlations between (reflectance, illumination), (table, tv-monitor), materials (wooden, plastic) given For example, model enforce condition that if set pixels take same label, e.g. table, most likely those would receive similar values. We cast problem energy minimization demonstrate qualitative quantitative improvement overall accuracy on NYU Pascal datasets.