Shape from Shading under changing Illumination using a CNN

This paper presents an alternative method to the classical yet foundational problem of recovering the geometry of a 3D object from a single 2D brightness image, i.e., shape-from-shading. Drawing inspiration from traditional approaches and recent work in single-image 3D reconstruction, we employ a convolutional neural network trained on synthetic images of characteristic shapes as well as high-frequency surfaces to address the shape-from-shading problem. Although the resulting network is less accurate than non-deep learning (DL) approaches, it exhibits greater resilience to common pitfalls, such as variations in light source direction and the presence of noise in the input image.