目标检测知识蒸馏综述OA

High-precision object detection models demand substantial computational resources and contain massive parame-ters,making their deployment on resource-constrained edge devices challenging.Knowledge distillation,as an efficient model compression technique,effectively transfers knowledge from a complex teacher object detector to a lightweight student object detector,maximizing the performance of the deployed detection model.Currently,there is a lack of comprehensive surveys on knowledge distillation for object detection.To bridge this gap,this study began by categorizing mainstream knowledge distil-lation for object detection methods into three major types:logits-based distillation,feature-based distillation,and relation-based distillation.Then it reviewed the research progress in each category,comparing and summarizing their respective advan-tages and disadvantages.Furthermore,it made performance comprehensive comparison and analysis of these three approaches on the MS COCO dataset and three mainstream detector architectures.Finally,this paper outlined future research directions and challenges about knowledge distillation for object detection method.