PBDE: An Effective Method for Filtering False Positive Boxes in Object Detection

摘要

An inevitable problem in the practice of object detection is the existence of false positive detection boxes. False detection boxes greatly reduce precision of object detection model and compromise the desired effect. In this paper, we propose a method named Prediction Box Density Evalution (PBDE). We summarize box density characteristics of true positive (TP) and false positive (FP) boxes to filter out a large number of FP boxes. After PBDE, we can obtain a significant improvement in precision with a low recall loss, and increase F1-score by up to 9 percentage when the confidence threshold is 0.1. The entire algorithm is carried out on the post-processing of object detection, and there is no need to change the original training method and network structure, which is of great practical importance.