search

med-air / Endo-FM

[MICCAI'23] Foundation Model for Endoscopy Video Analysis via Large-scale Self-supervised Pre-train
Apache License 2.0
146 stars 14 forks source link

Color normalization #11

Closed jmorlana closed 3 weeks ago

jmorlana commented 5 months ago

Hi, I have a question about the normalization in the pretraining. Did you use standard Imagenet values to normalize or did you normalize according to the medical data domain? Thanks!!

Kyfafyd commented 5 months ago

Hi, thanks for your interest! The normailzation is using the ImageNet values: https://github.com/med-air/Endo-FM/blob/c0979d2d235a61f85daf171a73a34f47683a67d9/datasets/transform.py#L679-L681

jmorlana commented 5 months ago

Hi, thanks for your reply and your amazing work! And did you try with custom normalization values or is something you did not considered? I would say medical image data is quite different from the usual natural images used in ImageNet.

Kyfafyd commented 5 months ago

Hi, from my prevous practice, the normalization value is not that important. But you can still try to use a specific normalization value to fit. If you have any findings, welcome to post here!