Segmentation of illuminated areas of scene using fully-convolutional neural networks and computer vision algorithms for augmented reality systems