Single Image Shadow Detection using Multiple Cues in a Supermodular MRF

In this paper we propose a complete methodology for single image shadow detection based on the learned appearance of shadows. The basis of our method is a novel single region SVM classifier with a multi-kernel model specifically tailored for shadow region classification. This classifier, which already outperforms much more complex methods, provides the unary potentials for an MRF optimization that also includes pairwise potentials encoding the relationships between neighboring regions in the image. We introduce a novel boundary classifier for shadow boundaries cast over surfaces with the same material, and two improved paired regions classifiers; one for adjacent regions of the same material taken under the same illumination, and one for regions of same material taken under different illumination. The strength of the unary classifier means that our MRF requires only relatively sparse pairwise potentials, resulting in a more efficient and accurate optimization as can be seen in our experimental results. We reduce the balanced error rate by 53% compared to the state of the art on the latest shadow detection image dataset.