Computer Vision News - September 2019

Research Every month, Computer Vision News reviews a research paper from our field. This month we have chosen Semantic Priors for Intrinsic Image Decomposition. We are indebted to the authors (Saurabh Saini and P. J. Narayanan) for allowing us to use their images. by Amnon Geifman Introduction Intrinsic Image Decomposition (IID) is the task of decomposing image into two parts; reflectance and shading. The reflectance component represents how the material of the object reflects light independent of viewpoint and illumination. The shading component captures the illumination effects caused by direct and indirect lighting in the scene. IID has many applications in computer vision and computer graphics such as contrast enhancement, image re-texturing, colorization and more. While the vision community has seen a significant advance in single image IID, it remains a challenging, highly ill-posed problem. Since the problem is ill-posed in nature, prior assumptions must be introduced. Reflectance sparsity is a common prior that assumes the color details are simple, i.e. it doesn't change much between nearby pixels. Shading smoothness is a prior which assumes the shading component of the image is a smooth surface. Optimizing these two priors separately might lead to inferior results. For this reason, the proposed method suggests a formulation to optimize these two priors in a single integrated algorithm using an iterative optimization. In order to better determine the sparse and the smooth regions, the authors incorporate semantic scene information from a DNN-based semantic segmentation. We next review the method in detail. 4