Towards Generalized UAV Object Detection: A Novel Perspective from Frequency Domain Disentanglement

When deploying unmanned aerial vehicle (UAV) object detection networks to complex, real-world scenes, generalization ability is often reduced due to domain shift. While most existing domain-generalized object detection methods disentangle domain-invariant features spatially, our exploratory experiments revealed a key insight for UAV object detection (UAV-OD): frequency domain contributions exhibit more pronounced disparities in generalization compared to generic object detection involving larger objects, since UAV-OD detects smaller objects. Therefore, frequency domain disentanglement stands out as a more direct, effective approach for UAV-OD. This paper proposes a novel frequency domain disentanglement method to improve UAV-OD generalization. Specifically, our framework leverages two learnable filters extracting domain-invariant and domain-specific spectra. Additionally, we design two contrastive losses: an image-level loss and an instance-level loss guiding training. These losses enable the filters to focus on extracting domain-invariant and domain-specific spectra, achieving better disentangling. Extensive experiments across multiple datasets, including UAVDT and Visdrone2019-DET, utilizing Faster R-CNN and YOLOv5, show our approach consistently and significantly outperforms baseline and state-of-the-art domain generalization methods. Our code is available at https://github.com/wangkunyu241/UAV-Frequency .