search

satoshi-kosugi / Unpaired-Image-Enhancement

54 stars 12 forks source link

Reproduce the results using generator_weights_for_demo.npz #11

Open ZZU-fjq opened 2 years ago

ZZU-fjq commented 2 years ago

Hello,

I tried to reproduce the result of paper by test.py and _generator_weights_for_demo.npz_.

I export images from fivek dataset, my setting for both raw(default/Input with Daylight WhiteBalance minus 1.5) and expertC as follow:

QQ图片20220418181104

And I use _generator_weights_fordemo.npz by adding _serializers.load_npz(args.loadgenerator, gen) in test.py. Nothing else has changed.

I calculated PSNR and SSIM by retouched image and target image. However, there is a big gap between the calculated results and the paper results.

Could you let me know which part is wrong.

Thank you very much.

satoshi-kosugi commented 2 years ago

As raw images, we used images that were not retouched at all. To create the raw images we used, please create a new catalog in Lightroom, import the all .dng files, and export them as they are.

If you want to use default/Input with Daylight WhiteBalance minus 1.5 as raw images, please retrain our model.