Repulsion Distance Intersection over Union Loss for Crowd Scene Detection

Abstract

The detection of many people or crowds increases the false-positive prediction rate owing to the creation of more bounding boxes. In this study, a loss function that reduces the false positive predictions of an object detection model is proposed. The proposed loss function induces not only the bounding box to be closer to the ground truth (GT) box but also the center point distance between the GT box and the overlapped adjacent object or other bounding boxes to be longer. In addition, a balanced feature pyramid was introduced to enhance the precision of object prediction. Applying the proposed method, the log-average miss rate on false positive per image in [10^(-2),100] was 1.89% lower while average precision and Jaccard index were 0.5% and 0.58% higher, respectively, than those of the conventional method, which indicates that it effectively reduces false positive predictions.