Boosting Domain Generalization in Object Detection through the Lens of Phase Invariance

Chenyang Zhang, Anqi Dong, Cong Hua, Yao Zhu, Lin Mei, Qing Cheng, Guangming Zhu, zhang liang

Keywords: Domain Generalization, Object Detection, Robust

TL;DR: We propose a Preserving Phase Invariance (PPI) paradigm, which significantly enhances model generalization in detection tasks under this constraint.

Abstract: Temporal and seasonal variations in dynamic real-world environments result in diverse visual appearances, posing significant challenges for object detection models to maintain consistently high performance. Although existing Domain Generalization (DG) methods have shown promise in enhancing model robustness, they often neglect the spatial structural relationships of objects during the learning of domain-invariant features, thereby limiting their effectiveness in object detection tasks compared to classification tasks. From the perspective of Preserving Phase Invariance (PPI), we propose a novel methodology that aims to enhance model generalization while preserving accurate object localization. This methodology comprises three complementary modules: Mix Normalization Perturbation (MNP), which synthesizes diverse styles to improve robustness; Sensitive Channel Perturbation (SCP), which suppresses domain-specific features at the channel level; and Attention on Amplitude (AOA), which applies spectral attention to the amplitude component. Together, these modules promote phase-invariant representations and contribute to improved cross-domain detection performance. Our approach fundamentally reduces the domain generalization gap in classification and detection by maintaining the integrity of key structural information. Our proposed methods achieve state-of-the-art performance on Unsupervised Domain Adaptation and Single Domain Generalization Object Detection benchmarks, even outperforming most recent state-of-the-art Domain Adaptation techniques. The code is available in the supplementary material.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 12509