Forums

NTIRE 2025 Real-World Face Restoration Forum

Go back to competition Back to thread list Post in this thread

> scores obtained from the local deployment of iqa-pytorch are different from the scores obtained upon submission.

Hello,
I sincerely appreciate your efforts for oorganize this competition.
I recently executed the iqa - pytorch project on my local system. To my surprise, the CLIPIQA scores I calculated locally deviate from those obtained from the website validation server.
Given your statement that "All metrics will be tested using the default settings of IQA - PyTorch," I'm left to wonder whether I might have made some errors during my local implementation.

Posted by: fusion @ Feb. 27, 2025, 6:28 a.m.

I also meet this issue. After I changed the image format from RGB to BGR, the CLIPIQA score got close to server's score, err is less than 0.01.

Posted by: changxin.zhou @ Feb. 27, 2025, 6:51 a.m.

thanks to changxin.zhou's response!

Posted by: fusion @ Feb. 27, 2025, 6:57 a.m.

The evaluation code in codalab is updated to RGB. Sincerely apologize for the potential inconvenience made by us. The final test will be conducted in RGB images, not BGR.

Posted by: jkwang @ Feb. 28, 2025, 7:32 a.m.
Post in this thread