Gaussian-Driven Unsupervised Domain Adaptation Object Detection Transformer for Remote Sensing Imagery

Domain shift can cause the detector to fail in adapting to the test data, resulting in sharp decline in detection performance. To address this issue, many unsupervised domain adaptation detection methods have been proposed to improve model's generalization ability, which are typically designed for natural scene data with simple backgrounds, where the intra-class features of the objects often follow a simple unimodal distribution. However, remote sensing images usually have complex background interference, exacerbating the domain shift. Additionally, the object class features in these images often exhibit multimodal distributions, making it difficult to effectively perform domain alignment within the same class. To overcome these challenges, we propose a DETR-based Gaussian-driven UDA method, named G-UDA. First, we introduce the real-time Gaussian frequency domain augmentation module, which increases the diversity of source domain data by perturbing the background information in specific frequency bands, while preserving object features. Next, we design the Gaussian class prototype alignment module based on Gaussian Mixture Model, which generates multiple Gaussian class prototypes, enabling the model to capture multimodal distribution within classes, thereby improving the representational capacity of intra-class features and enhancing inter-class discriminability. Finally, we integrate our foreground focus alignment module with the teacher-student self-training framework. By replacing the traditional global feature alignment with focus on precisely aligning features of foreground objects, we facilitate the extraction of more robust object features by the detector under complex background interference. Experimental results in multiple remote sensing domain adaptation scenarios demonstrate that our method outperforms current state-of-the-art methods.