search

zudi-lin / rcan-it

Revisiting RCAN: Improved Training for Image Super-Resolution
MIT License
91 stars 16 forks source link

Very interesting work! But there is a small question #2

Closed blackcow closed 2 years ago

blackcow commented 2 years ago
WX20220213-163454@2x

It seems that the Baseline is obtained w/o self-ensemble. However, the PSNR obtained with "Larger Patches" seems to be consistent with the RCAN-it+ (a self-ensemble version).

WX20220213-164957@2x

Is this a reasonable comparison? Or is there something wrong with my understanding, looking forward to your reply : )

zudi-lin commented 2 years ago

Hi @blackcow,

Thank you for your interest in our work!

  1. The Baseline is obtained with self-ensemble (Table 3 caption, "Results are produced with self-ensemble."). The Baseline is the same in Table 3 and Table 4. Code here: https://github.com/zudi-lin/rcan-it/blob/main/ptsr/model/arch.py#L197
  2. The final RCAN-it+ is indeed Update Baseline + SiLU & Longer Training + Larger Patches. So the evaluation scores are exactly the same.

Hope I have addressed your questions.

blackcow commented 2 years ago

All my questions are solved. Thank you very much for your reply !