compphoto / Intrinsic

What it does

We propose a method for generating high-resolution intrinsic image decompositions for in-the-wild images. Our method consists of multiple stages. We first estimate a grayscale shading layer using our ordinal shading pipeline. We then estimate low-resolution chromaticity information to account for colorful illumination effects while maintaining global consistency. Using this initial colorful decomposition, we estimate a high-resolution, sparse albedo layer. We show that our decomposition allows us to train a diffuse shading estimation network using only a single rendered indoor dataset.