Progressive Domain Adaptive Object Detection Based on Self‐Attention in Foggy Weather

Abstract

The quality of captured images is compromised by foggy weather, resulting in poor performance of detection networks. To mitigate this issue, researchers have attempted to employ domain adaptation techniques. However, a significant difference between the domains can create difficulties in the adaptation process, resulting in unstable training procedures and less‐than‐optimal outcomes. A self‐attention‐based progressive domain adaptive object detection framework is proposed to address this issue. We solved this problem by building an intermediate domain through physical synthesis. Moreover, the domain features of different levels are aligned in the feature space to achieve a better domain adaptation effect. In addition, we add a plug‐and‐play attention module to enhance the feature discriminability in the shallow feature extraction stage. Our method performs better than the existing state‐of‐the‐art techniques in synthetic and real‐world datasets, according to the experimental results. © 2023 Institute of Electrical Engineer of Japan and Wiley Periodicals LLC.