XVFI: eXtreme Video Frame Interpolation

Softmax Splatting for Video Frame Interpolation

“…(a) overlaid inputs (b) SoftSplat [44] (c) XVFI [57] (d) Ours Figure 8. Qualitative comparison of our proposed approach with two representative methods on a sample from the XTEST-2K [57] test dataset.…”

“…(a) overlaid inputs (b) SoftSplat [44] (c) XVFI [57] (d) Ours Figure 8. Qualitative comparison of our proposed approach with two representative methods on a sample from the XTEST-2K [57] test dataset. While these sophisticated interpolation methods are unable to handle this challenging scenario with the utility pole subject to large motion, our comparatively simple approach is able to generate a plausible result.…”

“…Figure 9. Evaluating the per-frame quality when performing 8× interpolation on XTEST-2K[57]. Our proposed approach not only generates better results, it also does so an order of magnitude faster thanks to our splatting-based synthesis technique which is particularly well-suited for interpolating multiple frames at once.…”

FILM: Frame Interpolation for Large Motion

“…Handling large motion is an important yet under-explored topic in frame interpolation. The work in [32] handles large motion by training on 4K sequences with extreme motion. While this is a viable approach, it does not generalize well on regular footage as discussed in [26].…”

“…Second, at test time, models perform well when the motion range matches that of the training datasets', but generalize poorly on more extreme motion. One could approach this with data that captures the desired range [32]. We instead propose a network that generalizes well in small and large motion.…”