Although the organizers emphasis the odd-numbered LR frames are not allowed to be used as input,
I think it is still not appropriate to release odd-numbered LR frames of the testset.
If the solution is designed to be a two-stage one: video frame interpolation + video super resolution,
how can we ensure that the odd-numbered frames of the testset are not used as supervision by participants to train the interpolation part (which is fed with even-numbered frames of the testset)?
It is also difficult to distinguish such a cheating method.
Posted by: hedlmote @ Jan. 14, 2021, 8:51 a.m.
Dear hedlmote,
We will do our best to make such circumstances would not happen.
As we have stated on the evaluation page, we set reproducibility as a necessary condition.
We will check the reproducibility by running the submitted source code on our own after the email submission.
No odd frames will be provided as input.
The corresponding frames in CodaLab-submitted frames, email-submitted frames, and our reproduced result should equal to each other. (unless there exists randomness in the algorithm itself)
Best,
Seungjun