Dual Prediction-Guided Distillation for Object Detection in Remote Sensing Images

Knowledge distillation, which intends to transfer the expertise from a complex teacher model to a concise student model, has achieved impressive success in object detection. However, many existing distillation methods are designed for object detection tasks on natural images and perform poorly on more challenging remote sensing images. In this study, we first attribute the deficiencies of knowledge distillation in remote sensing object detection to two core reasons: 1) lack of relation distillation among different instances and pixels and 2) significant differences in feature magnitude between the teacher-student pair. Then, we propose a dual prediction-guided knowledge distillation framework, which includes relation distillation and output distillation to address two issues, respectively. Prediction-guided relation distillation (PGRD) is proposed to capture the knowledge of global relation at both instance-wise and pixel-wise, then allowing students to better understand and depict the features distribution across different categories. Prediction-guided output distillation (PGOD) is proposed to mitigate the impact of feature magnitude inconsistencies on distillation with classification and location knowledge, then allowing students to directly capture task-relevant information. Finally, experimental results have indicated the consistent effectiveness of our method across anchor-based two-stage, one-stage, and anchor-free detectors with 11 comparison knowledge distillation methods on three remote sensing detection datasets. Source codes are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/RQ-W/DPGD.git</uri>.