Improving Semantic Segmentation via Efficient Self-Training