Learning Spatio-Temporal Downsampling for Effective Video Upscaling