An Adaptive Sampling Scheme to Efficiently Train Fully Convolutional Networks for Semantic Segmentation.