Normalization ¶
By: jsonneck on Feb. 29, 2024, 10:34 a.m.
Hi everyone, will the ground truth data be normalized in any way before the metrics are calculated?
By: dorian_kauffmann on March 14, 2024, 10:05 p.m.
Hi jsonneck,
Thanks for your patience,
Yes, a percentile normalization will be computed over each image before calculating the metrics.
We have updated the Database section (Test Database -> Normalization) :
https://lightmycells.grand-challenge.org/database/ with more information and the python code.
By: danifranco on March 23, 2024, 3:01 p.m.
Hi,
I think that there is something wrong in the piece of code you posted in Data section, as it gives you something strange. Look at the result applied to a Nucleus target image (left original/right percentile_normalization()
function applied with default settings): Image
I think that a clip should be done instead of the below line:
img_norm = dtype_max * (image - low_p) / ( high_p - low_p )
Is this result expected? Just to be sure...
Best,
Dani
By: pyatkovsky15022001 on March 28, 2024, 3:58 p.m.
Following danifranco post, I also encoutered unexpected results (https://ibb.co/6BjZsc9) after percentile_normalization()
.
Did we anticipate this outcome?
Sincerely, Sergey
By: dorian_kauffmann on March 29, 2024, 4:31 p.m.
Hello,
Thank you both of you for highlighting this !
After various checks, we chose to fix this error by deleting the "dtype_max" multiplication:
img_norm = (image - low_p) / ( high_p - low_p )
Then all the metrics are changed and more understandable.
So I've evaluate again all the submitted evaluation to get the updated metrics. The leaderboard changed a little with the rank order, sorry for that too.
We are sincerely sorry for the inconvenience!
If that can help you, we extend the Phase 1 until Thurday, April 04!
Sincerely,
Dorian