Work

PANDA: A Polarized Attention Network for Enhanced Unsupervised Domain Adaptation in Semantic Segmentation

Year: 2024

Type: article

Abstract: Unsupervised domain adaptation (UDA) focuses on transferring knowledge from the labeled source domain to the unlabeled target domain, reducing the costs of manual data labeling. The main challenge in UDA is bridging the substantial feature distribution gap between the source and target domains. To address this, we propose Polarized Attention Network Domain Adaptation (PANDA), a novel approach that leverages Polarized Self-Attention (PSA) to capture the intricate relationships between the source and target domains, effectively mitigating domain discrepancies. PANDA integrates both channel and spatial information, allowing it to capture detailed features and overall structures simultaneously. Our proposed method significantly outperforms current state-of-the-art unsupervised domain adaptation (UDA) techniques for semantic segmentation tasks. Specifically, it achieves a notable improvement in mean intersection over union (mIoU), with a 0.2% increase for the GTA→Cityscapes benchmark and a substantial 1.4% gain for the SYNTHIA→Cityscapes benchmark. As a result, our method attains mIoU scores of 76.1% and 68.7%, respectively, which reflect meaningful advancements in model accuracy and domain adaptation performance.

Source: Electronics

DOI: 10.3390/electronics13214302

Authors: Chiao-Wen Kao, Wei-Ling Chang, Chun-Chieh Lee, Kuo-Chin Fan

Institutions: Ming Chuan University, National Central University

Language: English

Topic: Domain Adaptation and Few-Shot Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,728

Funder: Ministry of Science and Technology, Taiwan