search

xuebinqin / DIS

This is the repo for our new project Highly Accurate Dichotomous Image Segmentation
Apache License 2.0
2.27k stars 270 forks source link

About preprocessing / normalizing #63

Closed faruknane closed 1 year ago

faruknane commented 1 year ago

Hello @xuebinqin ,

In u2net, if I am not wrong, you were using imagenet normalization. However in this paper, you are using normalize(image,[0.5,0.5,0.5],[1.0,1.0,1.0]) . Is there a particular reason to do this? Or can we also go with the normalization in the u2net?

Thank you so much!

xuebinqin commented 1 year ago

according to our experiences, the imagenet nromalization doesn't matter. Besides, our model in DIS is trained without using pre-trained imagenet weights.

On Thu, Dec 8, 2022 at 11:41 AM Akif Faruk Nane @.***> wrote:

Hello @xuebinqin https://github.com/xuebinqin ,

In u2net, if I am not wrong, you were using imagenet normalization. However in this paper, you are using normalize(image,[0.5,0.5,0.5],[1.0,1.0,1.0]) . Is there a particular reason to do this? Or can we also go with the normalization in the u2net?

Thank you so much!

— Reply to this email directly, view it on GitHub https://github.com/xuebinqin/DIS/issues/63, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORNVEXHFMRQ7QI46AQLWMI2UZANCNFSM6AAAAAASYQR2C4 . You are receiving this because you were mentioned.Message ID: @.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

faruknane commented 1 year ago

Okay, so we can also use imagenet normalization for DIS, right?

xuebinqin commented 1 year ago

when retrain the model, you can use different normalization strategies. For using our provided model weights, the normalization should be kept as it is now.

On Sat, Dec 10, 2022 at 3:16 PM Akif Faruk Nane @.***> wrote:

Okay, so we can also use imagenet normalization for DIS, right?

— Reply to this email directly, view it on GitHub https://github.com/xuebinqin/DIS/issues/63#issuecomment-1345406155, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADSGORJRIAMJUQTG2BIC7DDWMUFOFANCNFSM6AAAAAASYQR2C4 . You are receiving this because you were mentioned.Message ID: @.***>

-- Xuebin Qin PhD Department of Computing Science University of Alberta, Edmonton, AB, Canada Homepage: https://xuebinqin.github.io/

faruknane commented 1 year ago

Thank you so much! I was just wondering if anything was behind this choice. @xuebinqin