PhysAug: A Physical-guided and Frequency-based Data Augmentation for Single-Domain Generalized Object Detection

• DOI: 10.1609/aaai.v39i20.35487
• Publication Date: 2025-04-11T13:08:56Z
• Authors: Xiaoran Xu; Jiangang Yang; Wenhui Shi; Siyuan Ding; Luqing Luo; Jian Liu

Single-Domain Generalized Object Detection (S-DGOD) aims to train on a single source domain for robust performance across variety of unseen target domains by taking advantage an object detector. Existing S-DGOD approaches often rely data augmentation strategies, including composition visual transformations, enhance the detector's generalization ability. However, absence real-world prior knowledge hinders from contributing diversity training distributions. To address this issue, we propose PhysAug, novel physical model-based non-ideal imaging condition method, adaptability tasks. Drawing upon principles atmospheric optics, develop universal perturbation model that serves as foundation our proposed PhysAug. Given perturbations typically arise interaction light with particles, image frequency spectrum is harnessed simulate variations during training. This approach fosters detector learn domain-invariant representations, thereby enhancing its ability generalize various settings. Without altering network architecture or loss function, significantly outperforms state-of-the-art datasets. In particular, it achieves substantial improvement 7.3% and 7.2% over baseline DWD Cityscape-C, highlighting enhanced generalizability in