CPLM: correct pseudo-label of mixed-mode for domain adaptive semantic segmentation under adverse conditions

Pan Ouyang; Zhaoxiang Wang; Xiaoguo Yao

doi:10.1117/12.3025677

8 April 2024 CPLM: correct pseudo-label of mixed-mode for domain adaptive semantic segmentation under adverse conditions

Pan Ouyang, Zhaoxiang Wang, Xiaoguo Yao

Author Affiliations +

Proceedings Volume 13090, International Conference on Computer Application and Information Security (ICCAIS 2023); 1309012 (2024) https://doi.org/10.1117/12.3025677
Event: International Conference on Computer Application and Information Security (ICCAIS 2023), 2023, Wuhan, China

Abstract

Self-training is a popular approach for domain adaptation that relies on generating high-quality pseudo labels. However, due to domain shift, pseudo labels inevitably contain noise. Existing methods usually consider target domain entropy regularization single to correct pseudo label noise. Alternatively, some methods consider how to better mix the source and target domains to mitigate unexpected minimization of pseudo label entropy. But in the mixed mode, the pseudo label presents new uncertainty characteristics and still has space for correction. Therefore, this paper proposes a method that further considers uncertainty of target domain predictions and the use of dynamic thresholding to correct the pseudo label in the mixed domain for better knowledge transfer. Specifically, the proposed pseudo-label correction method uses different loss weights for the mixed domain according to the uncertainty, where the weight of the target domain is determined according to both local uncertainty and overall proportion greater than a certain threshold. Furthermore, a dynamic threshold method is used according to different stages of training and different classes. Our approach achieves competitive results on three adaptive semantic segmentation tasks in Cityscapes→ACDC, Cityscapes→FoggyCityscapes +RainCityscapes, Cityscapes→Dark Zurich in normal to adverse scenes.

(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Pan Ouyang, Zhaoxiang Wang, and Xiaoguo Yao "CPLM: correct pseudo-label of mixed-mode for domain adaptive semantic segmentation under adverse conditions", Proc. SPIE 13090, International Conference on Computer Application and Information Security (ICCAIS 2023), 1309012 (8 April 2024); https://doi.org/10.1117/12.3025677

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available

Members: $17.00

Non-members: $21.00 ADD TO CART

PROCEEDINGS
11 PAGES

DOWNLOAD PAPER SAVE TO MY LIBRARY

GET CITATION

RIGHTS & PERMISSIONS

Get copyright permission Get copyright permission on Copyright Marketplace

KEYWORDS

Education and training

Semantics

Image segmentation

Performance modeling

Adversarial training

Data modeling

Transformers

Show All Keywords

Keywords/Phrases

Search In:

Publication Years