search

XY-boy / EDiffSR

[IEEE TGRS 2024] EDiffSR: An Efficient Diffusion Probabilistic Model for Remote Sensing Image Super-Resolution
103 stars 6 forks source link

Metrics calculation #8

Open aleksmirosh opened 5 months ago

aleksmirosh commented 5 months ago

Hello, thank you for this really interesting project for remote sensing. I fond that the script for testing is not complete. I tried to test checkpoints with AID, btw thank you for providing the weights. But I got a different result. Could you please add metrics calculation?

Like I got SSIM 0.2 on AID test.

XY-boy commented 5 months ago

Hello, thank you for this really interesting project for remote sensing. I fond that the script for testing is not complete. I tried to test checkpoints with AID, btw thank you for providing the weights. But I got a different result. Could you please add metrics calculation?

Like I got SSIM 0.2 on AID test.

SSIM=0.2, this result should be problematic. Does the visual result look correct? btw, are you testing with aid.yml, sorry the settings for the model parameters in aid.yml are not the same as the pre-trained model, I've updated it. You can also use nwpu.yml for testing.

aleksmirosh commented 5 months ago

SSIM=0.2, this result should be problematic. Does the visual result look correct? btw, are you testing with aid.yml, sorry the settings for the model parameters in aid.yml are not the same as the pre-trained model, I've updated it. You can also use nwpu.yml for testing.

Thank you for your answer. I tested nwpu.yml on NWPU dataset. I got ssim around 0.17 I randomly selected 315 images, center cut to 128x128, and then interpolated with bicubic to 512x512. I use 128x128 as patches as LR and interpolated as GT. Is this process correct?

The results look very noisy but promising.

XY-boy commented 5 months ago

SSIM=0.2, this result should be problematic. Does the visual result look correct? btw, are you testing with aid.yml, sorry the settings for the model parameters in aid.yml are not the same as the pre-trained model, I've updated it. You can also use nwpu.yml for testing.

Thank you for your answer. I tested nwpu.yml on NWPU dataset. I got ssim around 0.17 I randomly selected 315 images, center cut to 128x128, and then interpolated with bicubic to 512x512. I use 128x128 as patches as LR and interpolated as GT. Is this process correct?

The results look very noisy but promising.

Hi. You can handle it as follows:

  1. use center cropping to crop the 512*512 image from the randomly selected NUPU images as the ground truth (GT)
  2. use matlab's imresize function to downsample the 512512 GT images to 128128 low-resolution versions (LR) by bicubic interpolation
  3. Use EDiffSR to super-resolve the LR to 512*512 SR results, and calculate SSIM between SR and GT.
aleksmirosh commented 5 months ago

@XY-boy thank you, it goes up to 0.35 for SSIM. The images look noisy.