Original title: Intrinsic Image Decomposition via Ordinal Shading
Authors: Chris Careaga, Yağız Aksoy
This work delves into the challenge of intrinsic decomposition, pivotal in computational photography. The process entails precisely estimating shading and albedo. Their approach involves two phases: creating a dense ordinal shading model using a shift- and scale-invariant approach, followed by combining low- and high-resolution estimates for a coherent shading map. The model is incentivized to learn accurately through specific losses on shading and albedo. They propose a method to create pseudo ground truth from model predictions and multi-illumination data, ensuring adaptability to real-world images. Through extensive analysis, their method shows superior results compared to existing techniques. Notably, it enables complex image edits like recoloring and relighting that were previously challenging. This research showcases advancements in accurately dissecting intrinsic components, crucial for enhancing computational imagery techniques.
Original article: https://arxiv.org/abs/2311.12792