Clarification on PSNR/SSIM Calculation and Dataset Use for DRCT-L Model on Set5

[Ar0Kim]

Hello,

Thank you for sharing your project. I have a question regarding the PSNR and SSIM results for the DRCT-L model. Specifically, for a dataset like Set5, are the PSNR and SSIM values reported in the paper calculated as the average over each super-resolved image's individual PSNR and SSIM scores?
When I run the test.py file using the DRCT-L_X4.pth model, the PSNR log only shows a single best score, as seen below.

Did you use a separate script to calculate the PSNR and SSIM for each image individually and then compute their average?

Additionally, I am trying to replicate the paper’s result for Set5 inference with the DRCT-L model, which reports a PSNR of 33.37 and an SSIM of 0.9090. I downloaded the dataset from the link provided (Validation data link, https://github.com/ChaofWang/Awesome-Super-Resolution/blob/master/dataset.md),

and for Set5, the folder structure includes images ending in “LR.png” and “HR.png”.

In this context, should the dataroot_lq point to the folder containing images ending with “LR.png” and dataroot_gt point to the folder containing images ending with “HR.png”? If not, please let me know which PNG files should be used as HR and LQ images.

To achieve the reported PSNR and SSIM values, could you please clarify which exact dataset was used and provide any details or code regarding the metric calculations? Currently, the best scores I obtain are lower than those presented in the paper.

Thank you very much for your assistance.

[riccardolagrassa]

Hello,

Thank you for sharing your project. I have a question regarding the PSNR and SSIM results for the DRCT-L model. Specifically, for a dataset like Set5, are the PSNR and SSIM values reported in the paper calculated as the average over each super-resolved image's individual PSNR and SSIM scores? When I run the test.py file using the DRCT-L_X4.pth model, the PSNR log only shows a single best score, as seen below.

Did you use a separate script to calculate the PSNR and SSIM for each image individually and then compute their average?

Additionally, I am trying to replicate the paper’s result for Set5 inference with the DRCT-L model, which reports a PSNR of 33.37 and an SSIM of 0.9090. I downloaded the dataset from the link provided (Validation data link, https://github.com/ChaofWang/Awesome-Super-Resolution/blob/master/dataset.md), and for Set5, the folder structure includes images ending in “LR.png” and “HR.png”.

In this context, should the dataroot_lq point to the folder containing images ending with “LR.png” and dataroot_gt point to the folder containing images ending with “HR.png”? If not, please let me know which PNG files should be used as HR and LQ images.

To achieve the reported PSNR and SSIM values, could you please clarify which exact dataset was used and provide any details or code regarding the metric calculations? Currently, the best scores I obtain are lower than those presented in the paper.

Thank you very much for your assistance.

Using the same model (finetuned for 250k iteraction over DF2K) i achieved the following results:
PSNR average: 31.1761474609375
SSIM average: 0.877326250076294

The results are not aligned with the paper, can you explain please?

[tututututurana]

你好

感谢您分享您的项目。我对 DRCT-L 模型的 PSNR 和 SSIM 结果有疑问。具体来说，对于像 Set5 这样的数据集，论文中报告的 PSNR 和 SSIM 值是否计算为每个超分辨率图像的单个 PSNR 和 SSIM 分数的平均值？当我使用 DRCT-L_X4.pth 模型运行 test.py 文件时，PSNR 日志仅显示一个最佳分数，如下所示。

您是否使用单独的脚本单独计算每张图像的 PSNR 和 SSIM，然后计算它们的平均值？

此外，我正在尝试使用 DRCT-L 模型复制论文的 Set5 推理结果，该模型报告的 PSNR 为 33.37，SSIM 为 0.9090。我从提供的链接（验证数据链接，https://github.com/ChaofWang/Awesome-Super-Resolution/blob/master/dataset.md）下载了数据集，对于 Set5，文件夹结构包括以“LR.png”和“HR.png”结尾的图像。

在这种情况下，dataroot_lq是否应该指向包含以“LR.png”结尾的图像的文件夹，dataroot_gt指向包含以“HR.png”结尾的图像的文件夹？如果没有，请让我知道哪些 PNG 文件应该用作 HR 和 LQ 图像。

为了获得报告的 PSNR 和 SSIM 值，您能否澄清使用了哪个确切的数据集，并提供有关指标计算的任何详细信息或代码？目前，我获得的最佳分数低于论文中介绍的分数。

非常感谢您的帮助。

I'm having the same problem. Did you find a solution?

[Stefan-084]

same question, could the author clarificate this?