Scale-aware spatial pyramid pooling with both encoder-mask and scale-attention for semantic segmentation.