Unsupervised Domain Adaptation for One-Stage Detector in Remote Sensing Imagery
Sihao Luo, Li Ma, Xingmei Li · 2023
We propose a novel domain-adaptive one-stage framework to address the unsupervised domain adaptation for object detection in remote sensing. To benefit from the inherent multiscale representations of the one-stage detector, several domain discriminators are used at different scales to help the model generate domain-invariant features for objects of different sizes. Moreover, we implicitly learn the invariant features from image-level/global to instance-level/local by exploiting a self-attention mechanism that allows the network to gradually recognize local regions that are crucial for detection and adaptation. The experimental results on various remote sensing datasets confirm the effectiveness of the proposed framework. Compared to the baseline model trained on the source dataset, our method consistently improves the detection performance on the target dataset by 4.5%-11.6% mAP under two domain adaptation scenarios and achieves new state-of-the-art performance.