voxel-flow training datasets

Hello. How to build the training dataset from UCF?

Feb 07 '18 02:02 qingsenyangit

I have a fork at https://git.sesse.net/?p=voxel-flow which contains, among others, a script to extract usable frames. I doubt it's exactly the same as the paper used (I assume that was some sort of MapReduce), but it allowed me to at least train a network with semi-reasonable results, so it can't be all bad.

Feb 10 '18 18:02 sesse

@sesse Could you tell me the meaning of "${PREFIX}.d1.psnr" in https://git.sesse.net/?p=voxel-flow ? Thank you!

Mar 03 '18 04:03 black107

The PSNR files are used to filter away frames without obvious motion; the dataset contains a lot of files that seem to have been upconverted from different frame rates. Basically I delay the file one frame and use ffmpeg to ask for PSNR between the two sequences, and if it's > 30 dB at any given spot (indicating a very good match), I can't use the triplet. Similarly, I delay it one more frame and check the PSNR again.

Mar 03 '18 06:03 sesse

@sesse Thanks for your comment. Could you tell me how do you get your training set? I download UCF dataset and randomly extract three consistent frames. Is this a proper way to generate the triplet? Or I misunderstand the meaning? Cause the result from this model is nothing near the one in the paper. The PSNR is only about 22 dB in cross validation and even worse in test set.