search

ceciliavision / zoom-learn-zoom

computational zoom from raw sensor data
http://people.eecs.berkeley.edu/~cecilia77/project-pages/zoom.html
320 stars 55 forks source link

About PSNR, SSIM and LPIPS on the released model #32

Open SSRheart opened 3 years ago

SSRheart commented 3 years ago

I have try many methods to reproduce the metrics on the released model. But those are still different with the metrics on the paper.

Does someone can give me some advise? Thanks in advance~

YuqiangY commented 3 years ago

Have you ever aligned the RGB data before calculating the metrics？ I calculate the PSNR/SSIM after data alignment, the metrics on the released model are 26.58/0.775. However, the metrics of the bicubic method are 25.63/0.736, which is much higher than the value in the paper. I am very confused about it. So，how do you reproduce the metrics?

SSRheart commented 3 years ago

Have you ever aligned the RGB data before calculating the metrics？ I calculate the PSNR/SSIM after data alignment, the metrics on the released model are 26.58/0.775. However, the metrics of the bicubic method are 25.63/0.736, which is much higher than the value in the paper. I am very confused about it. So，how do you reproduce the metrics?

I have aligned the HR groundtruth to the SR results(and the bicubic results) before calculate the metrics. but the metrics are still lower than the values in the paper. Then I find the histogram matching codes in the commits. After applying the histogram matching and alignment operation, I achieve the same situation with yours.