Efficient Semantic Segmentation by Altering Resolutions for Compressed Videos