Intrinsic Image Decomposition

Intrinsic image decomposition aims to separate a photograph into its constituent parts: surface reflectance (albedo) and illumination (shading), a challenging inverse problem due to its inherent ambiguity. Recent research focuses on improving decomposition accuracy using deep learning models, often incorporating physical constraints like light transport models or leveraging additional data sources such as LiDAR intensity or multi-view videos to overcome the ill-posed nature of the problem. These advancements enable applications in image editing (relighting, specularity removal), photogrammetry (albedo recovery for enhanced 3D modeling), and hyperspectral image analysis (improved classification). The field is actively exploring unsupervised and self-supervised learning approaches to reduce reliance on scarce ground truth data.