A Review of Knowledge Distillation in Object Detection

Abstract

Target detection is a revolutionary advancement in computer vision that provides the ability to identify specific targets in images for a wide variety of applications, including but not limited to video surveillance, face recognition, and autonomous driving. Although target detection has been developed to a high level and can be deployed for applications in several fields, there are still some problems in practice, such as the two-phase detection algorithm has high detection accuracy but slow detection speed, while the one-phase detection algorithm is fast, but its accuracy is poor. We need to combine their respective advantages further for related algorithm research, and we need to reach a balance between detection speed and detection accuracy so that the algorithms can be deployed on edge devices with limited computational power. Knowledge Refinement, as one of the common means of model compression, can solve the above problems, and it reduces the deployment difficulty of the algorithms. In this paper, we summarize the use of knowledge compression on target detection. We conclude the methods mentioned in the paper from an objective and unbiased perspective and suggest possible improvement directions, and we provide an outlook on the future trend of combining distillation learning and target detection. This paper provides a clear overview of the field of target detection and provides an idea of future trends.