Causal Distillation Prompt for Few-Shot Object Detection in Remote Sensing Images

Few-shot object detection (FSOD) recently garnered increasing attention in remote sensing images (RSIs). Although FSOD has achieved success in natural sense, the large size and complex backgrounds of RSIs make FSOD especially challenging. Recently, vision–language models have achieved great success in the computer vision field, offering new paradigms for FSOD in RSIs. This article proposes a causal distillation prompt for few-shot object detection in remote sensing images (CDP-FSDet) by learning efficient features for novel classes. First, a structural causal model tailored for FSOD in RSIs is developed to systematically analyze empirical errors in the knowledge distillation process. Second, the class-disentangle knowledge distillation is proposed to eliminate the empirical errors by implementing conditional causal interventions, ensuring more accurate and robust knowledge transfer between base and novel classes. Third, prompt adapt generation is introduced to generate prompts specifically tailored for RSIs, addressing the complex backgrounds and diverse object scales that differ from natural images. Extensive experiments on the DIOR and NWPU VHR-10.v2 datasets demonstrate that the proposed CDP-FSDet method significantly outperforms several state-of-the-art methods.