Kernel-based learning of cast shadows from a physical model of light sources and surfaces for low-level segmentation