Zooming Slow-Mo: Fast and Accurate One-Stage Space-Time Video Super-Resolution

Overlayed LR inputs HR intermediate frame Overlayed LR inputs DAIN+Bicubic DAIN+EDVR Ours Figure 1: Example of space-time video super-resolution. We propose a one-stage space-time video super-resolution (STVSR) network to directly predict high frame rate (HFR) and high-resolution (HR) frames from the corresponding lowresolution (LR) and low frame rate (LFR) frames without explicitly interpolating intermediate LR frames. A HR intermediate frame t and its neighboring low-resolution frames: t -1 and t + 1 as an overlayed image are shown. Compare to a state-ofthe-art two-stage method: DAIN [1]+EDVR [37] on the HR intermediate frame t, our method is more capable of handling visual motions and therefore restores more accurate image structures and sharper edges. In addition, our network is more than 3 times faster on inference speed with a 4 times smaller model size than the DAIN+EDVR.