ShadowNet

Abstract

Finding shadows in images is useful for many applications, such as white balance, shadow removal, or obstacle detection for autonomous vehicles. Shadow segmentation has been investigated both by classical computer vision and machine learning methods. In this paper, we propose a simple Convolutional-Neural-Net (CNN) running on a PC-GPU to semantically segment shadowed regions in an image. To this end, we generated a synthetic set of shadow objects, which we projected onto hundreds of shadow-less images in order to create a labeled training set. Furthermore, we suggest a novel loss function that can be tuned to balance runtime and accuracy. We argue that the combination of a synthetic training set, a simple CNN model, and loss function designed for semantic segmentation, are sufficient for semantic segmentation of shadows, especially in outdoor scenes.