Cross-Domain Remote Sensing Image Object Detection Based on Multi-Scale Domain Adaptive Teacher Network

Cross-domain remote sensing image object detection is highly dependent on the annotation of data and can be difficult to label. In addressing this problem, numerous researchers have proposed the utilization of pseudo-labeling approaches to mitigate domain shift, especially in two-stage detectors, which have shown significant effectiveness. However, the application of using a classical mean teacher network to generate pseudo-labels in single-stage detection models poses challenges. To address this problem, we propose a single-stage teacher model based on adversarial learning. Specifically, the teacher network is divided into classification and location sub-task branches, with an emphasis on optimizing pseudo-labeling. Compared to the mean teacher model, we pay more attention to the reliability of the boundary box. In the student network, this paper uses a multi-scale domain adaptive mechanism to extract domain invariant features of cross-domain images and uses an exponential sliding average to update the teacher network parameters, thereby improving pseudo labels and further narrowing the distribution differences. Experimental results confirm that compared to other single-stage detection methods for cross-domain object detection, this paper showcases a superior performance improvement in the field of remote sensing images.