search

psychopa4 / PFNL

Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations
MIT License
138 stars 21 forks source link

Some things about the results presented in the corresponding paper #7

Open ShihuaHuang95 opened 4 years ago

ShihuaHuang95 commented 4 years ago

Hi, this is a great job! I have some confusions about this paper: Table 3 shows the results over UDM10 dataset. I can not achieve such higher results (average about 28+ and 0.84+ for PSNR and SSIM, respectively), though our model which trained over Vimeo90K is outperformed all SOTA methods except yours when evaluated over Vid4. In the last, could you please provide the codes for PSNR and SSIM calculation?

Looking for your reply! Many thanks for your great work!

ShihuaHuang95 commented 4 years ago

One more thing, "Further, we have collected another 20 video sequences for evaluation during training" in Part 4.1 Implementation Details, what does it mean "for evaluation during training"? Does the results presented in 4.2, 4.3 and 4.4 are achieved when tested over such val dataset?

psychopa4 commented 4 years ago

Hi, this is a great job! I have some confusions about this paper: Table 3 shows the results over UDM10 dataset. I can not achieve such higher results (average about 28+ and 0.84+ for PSNR and SSIM, respectively), though our model which trained over Vimeo90K is outperformed all SOTA methods except yours when evaluated over Vid4. In the last, could you please provide the codes for PSNR and SSIM calculation?

Looking for your reply! Many thanks for your great work!

Models shown in the paper are all trained over MM522 dataset, you may as well try to train your models on this dataset. PSNR and SSIM are computed using Matlab

psychopa4 commented 4 years ago

One more thing, "Further, we have collected another 20 video sequences for evaluation during training" in Part 4.1 Implementation Details, what does it mean "for evaluation during training"? Does the results presented in 4.2, 4.3 and 4.4 are achieved when tested over such val dataset?

It means we use another 20 video sequences to evaluate the performance of the model during training. Yes, they are.

ShihuaHuang95 commented 4 years ago

Many thanks for the provided metrics!

wenchen4321 commented 4 years ago

Hi, this is a great job! I have some confusions about this paper: Table 3 shows the results over UDM10 dataset. I can not achieve such higher results (average about 28+ and 0.84+ for PSNR and SSIM, respectively), though our model which trained over Vimeo90K is outperformed all SOTA methods except yours when evaluated over Vid4. In the last, could you please provide the codes for PSNR and SSIM calculation?

What downsampling method of your training dataset Vimeo90K? I trained the model with Vimeo90K with bicubic downsampling, the results are very bad.

suxi-du commented 3 years ago

can u release your psnr and ssim evaluation code in matlab? The link is losing its efficacy. I got much worse result according to my understanding. it's just 20db.

psychopa4 commented 3 years ago

can u release your psnr and ssim evaluation code in matlab? The link is losing its efficacy. I got much worse result according to my understanding. it's just 20db.

I upload the PSNR and SSIM calculation code in Matlab, please see ./matlab/ .