Relation Perception Distillation for Object Detection

Abstract

Knowledge distillation is of great significance for the lightweight of object detectors. In this paper, we start with the issue of inharmonious predictions for object detectors and propose a novel Relation Perception Distillation (RPD), which consists of Correlation Distillation (CD) and Feature Perception Distillation (FPD). CD intuitively improves the harmony of predictions by inculcating the student model with correlation knowledge of classification scores and localization quality. Moreover, we observe that detector making inconsistent predictions is distracted by regions where features are mixed. Therefore, FPD is further proposed to assist detector in distinguishing features from different ground truths (GTs). Extensive experiments demonstrate the effectiveness and robustness of the proposed method. The results show that our method achieves up to 4.0 AP improvement on COCO2017 dataset, outperforming the recent Knowledge Distillation methods like FGD, TBD and DiffKD.

Keywords: object detection, knowledge distillation, computer vision