Testing scripts of VideoINR about videoinr-continuous-space-time-super-resolution

[sichun233746]

@hzwer Actually, I get little lower y-channel PSNR than table1.

I think you should only calculate PSNR for 1st to 8th frame, since the 9th frame will be the 1st frame in the next sequence.
Did you calculate PSNR value like this?

[hzwer]

@sichun233746 Thank you so much, I think it's reasonable. I will contact the author to check this.

[hzwer]

Thank you for this important suggestion again @sichun233746 , I have tested about 1/3 gopro images yet and got very close numbers as table 1.

[sichun233746]

Hello @zhangxydlut ,

Did you get the performance similar to the numbers in VideoINR paper ?
I'm trying to reproduce those numbers.
Can we discuss?

[hzwer]

Do you have any experience in reproducing Table 1 of the paper?

[sichun233746]

Do you have any experience in reproducing Table 1 of the paper?

@hzwer Yes, I've successfully re-produced the results.

[hzwer]

@sichun233746 I test the pretrained models, but get little higher y-channel PSNR than table 1 (Both Adobe240fps and Gopro). When calculating on rgb space. get about 1dB lower. Would you mind giving me some hints?

I follow the description of papre, "image sequences extracted from videos in the datasets are split into groups of 9-frame video clips. We feed the 1st and 9th frames down-sampled by scale ×4 in each clip into models to generate 9 high-resolution frames from 1st to 9th."
Paper Table 1 Gopro: 30.26dB, 29.41dB
My Gopro test: 30.95dB, 29.77dB (y channel)
My Gopro test: 28.2dB, 29.2dB (rgb channel)

[sichun233746]

@hzwer FYI, the Y channel PSNR I got for Gopro dataset is 30.08.