Interpolation results on UCF101 of DVF, SepConv, SuperSloMo trained on Adobe240fps, and SuperSloMo trained on YouTube240fps and Adobe240fps are provided here.

I used an incorrect version of motion masks, which were converted from motion mask provided by DVF. But somehow I overwrote the original motion masks. Since the motion masks I used for the paper look like dilated version. To reproduce the results in Table 4 of the CVPR 2018 paper, simply run "sh run get_results_paper.sh". You should be able to get similar results. (I regenerated results for SepConv and SuperSloMo_Adobe240fps. So the metrics are slightly different.)

I encourage you to use results generated by "sh get_results_bug_fixed.sh", which uses correct motion masks provided by DVF. Though different numbers are obtained from Table 4 in the CVPR 2018 paper, the relative order of different methods doesn't change and thus doesn't affect the conclusion of our paper. Other results reported in the paper don't use motion masks. Therefore, they are not affected by this bug. I'll update the arXiv paper after CVPR deadline.

There is another way of using motion masks, which is simply to "mask out" pixels outside of motion masks. It over estimates the accuracy of an interpolation algorithm, since for the regions outside the motion masks, it assumes the interpolation results are perfect (exactly the same pixel values between ground-truth and interpolation results).

Please feel free to contact Huaizu Jiang (hzjiang@cs.umass.edu) for any questions.