Layer-Wise Hierarchical Feature Transfer for Object Detection

In the realm of automatic driving, security monitoring, medical detection and other fields, object detection is widely used due to its remarkable object recognition and localization capabilities. Under this condition, integrating knowledge distillation technology into object detection has effectively alleviated the deployment challenges encountered in real-world scenarios. However, there exists a mismatch in the scale of teacher-student models, which causes a discrepancy in the magnitude of their prediction results. we propose a method called Layer-Wise Hierarchical Feature Transfer, which makes the student network learn at a more fine-grained level. Our method transfers student head's hierarchical features to the teacher's, and then fuses multiple results to replace the direct output by the student model for supervised loss computation. We conducted experiments using the MS COCO dataset. As shown by the results, our method achieves superior performance.