Numerical Results

When applied to LTR videos with severe motion blur and motion aliasing, frame interpolation methods (e.g., Nvidia SlowMo [9] and DAIN [2]) score significantly lower. However, even methods trained to overcome such challenges, but were trained on external datasets (Flawless [10]), struggle on videos that do not represent the typical motions and dynamic behaviors they were trained on. Videos 1-13 are such challenging examples.

Comparing Temporal Upsampling x8 Results on WAIC TSR Dataset

Detailed Comparison of Temporal Upsampling x8 Results on WAIC TSR Dataset

We compared average per-frame PSNR, SSIM and LPIPS[25] values of each method, as listed in the table above. To avoid boundary effects we did not include the first and last 30 frames of each sequence.We also disregarded a 20-pixel boundary around each frame when computing per-frame PSNR. This wide masking of the boundaries was done to accommodate large margin that some of the other algorithms require

The results above indicate that sophisticated frame-interpolation methods (DAIN [2], NVIDIA SloMo [9]) are not adequate for the task of Temporal Super Resolution (TSR), and are significantly inferior (-1 dB) on LTR videos compared to dedicated TSR methods (Ours and Flawless [10]). Flawless and Ours provide comparable quantitative results on the dataset, even though Flawless is a pre-trained supervised method, whereas Ours is unsupervised and requires no prior training examples. Moreover, on the subset of extremely challenging videos (with highly complex non-rigid motions), our Zero-Shot TSR outperforms the state-of-the-art externally trained Flawless[10].