DUDA: Online-Offline Dual Domain Adaption for Semantic Segmentation


Antao Pan (Zhejiang University),* Yawei Luo (Zhejiang University), Yi Yang (Zhejiang University), Jun Xiao (Zhejiang University)
The 33rd British Machine Vision Conference

Abstract

Self-training-based methods have achieved superior performance on unsupervised domain adaptive semantic segmentation task. However, these methods severely suffer from noisy pseudo label assignment. In this paper, we propose a simple yet effective dual pseudo label updating method that employs both online and offline mechanisms to dynamically update the two groups of pseudo labels. The online updating module employs a mean model to generate pseudo labels on-the-fly while the offline updating module capitalizes on the temporal consistency information to correct noisy labels. Furthermore, we present an online-offline dual regularization to further improve the noise-tolerant ability of the model. Combining the online-offline dual updating and online-offline dual regularization, we propose a novel mean-teacher based framework dubbed \textit{Online-Offline Dual Domain Adaption} (DUDA). Experiments show the proposed DUDA brings large performance gain and achieves state-of-the-art performance on two challenging benchmarks: GTA-to-Cityscapes and SYNTHIA-to-Cityscapes ($58.4\%$ mIoU and $59.7\%$ mIoU respectively).

Video



Citation

@inproceedings{Pan_2022_BMVC,
author    = {Antao Pan and Yawei Luo and Yi Yang and Jun Xiao},
title     = {DUDA: Online-Offline Dual Domain Adaption for Semantic Segmentation},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year      = {2022},
url       = {https://bmvc2022.mpi-inf.mpg.de/0585.pdf}
}


Copyright © 2022 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection