search

microsoft / ProDA

Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation (CVPR 2021)
https://arxiv.org/abs/2101.10979
MIT License
286 stars 44 forks source link

About the function full2weak() and label_strong_T() #40

Closed DeveloperZWK closed 1 year ago

DeveloperZWK commented 3 years ago

Dear authors: Thanks for sharing your code with us. I do not understand why you use augmentation methods to the output probability of the encoder in the function full2weak() and label_strong_T(). How do these two functions help? I sincerely hope you can help me:>.

JunXieFront commented 2 years ago

Dear authors: Thanks for sharing your code with us. I do not understand why you use augmentation methods to the output probability of the encoder in the function full2weak() and label_strong_T(). How do these two functions help? I sincerely hope you can help me:>.

Hi, I've got the same question, have you figured it out yet?

panzhang0104 commented 2 years ago

@DeveloperZWK @JunXieFront， we use full2weak for the output probability of the EMA encoder, because the input of EMA is the full image, while the input of the basic model is weak augmentation. In order to use information from EMA, we need to use full2weak() function to align two outputs.

label_strong_T() is a function to align the pseudo label with the input image. The strong augmentation includes some spatial deformations, and we need to apply the same spatial deformations to the supervision.

DeveloperZWK commented 2 years ago

您好，邮件已收，谢谢！

DeveloperZWK commented 2 years ago

您好，邮件已收，谢谢！

DeveloperZWK commented 1 year ago

您好，邮件已收，谢谢！