Multi-level domain perturbation for source-free object detection in remote sensing images

Recent advancements in cross-domain object detection have primarily relied on unsupervised domain adaptation (UDA) techniques to bridge domain gaps in remote sensing images. However, traditional UDA approaches require access to source domain data, creating privacy and transmission barriers. In response, we first introduce a novel setting called source-free object detection (SFOD) for remote sensing images. SFOD exclusively leverages pre-trained models from the source domain and unlabeled data from the target domain for adaptation. Our method emphasizes the detection model's focus on domain-invariant features by ensuring consistency between the perturbed and target domains. We initiate domain perturbation by manipulating domain-specific features at both image and feature levels, guided by considerations of style and color biases. Subsequently, we employ a multi-level alignment mechanism to ensure label-level and feature-level consistency between the perturbed and target domains. Our method incorporates the Mean Teachers framework and prototype-based feature distillation to alleviate noise arising from pseudo-labels. Extensive experiments conducted across various adaptation scenarios validate the effectiveness of our approach.