search

c-yn / FocalNet

[ICCV23] Focal Network for Image Restoration
MIT License
50 stars 3 forks source link

NH-Haze dataset #4

Closed chihlingchang closed 8 months ago

chihlingchang commented 8 months ago

Hi, I’m currently trying to train the model on NH-Haze dataset for image dehazing with the given code, and I had followed the settings showed in the Supplementary Material including setting training epoch=5000, batch size=2, initial learning rate=2e-4, cropping the images into 800*1200, and splitting the dataset into 45 training data, 5 validation data and 5 testing data.

However, my testing result (PSNR=19.78) is not as good as the performance shows in the paper (PSNR=20.43). I’d like to confirm whether there is any setting that I misunderstood, leading to the difference between our results?

Thank you so much!!

c-yn commented 8 months ago

Thanks for the question. We use 50 (45 training+5 validation) paired images for training following DeHamer.

chihlingchang commented 8 months ago

Thanks for your reply!

chihlingchang commented 8 months ago

Hi, I had tried to add the 5 validation paired images into training dataset, however, my testing PSNR is 20.06. I’m wondering if there is any other skills that I had missed can help to enhance the performance? Thank you!!