Object Detection Algorithm for Improving Non-Maximum Suppression Using GIoU
Zhiqiang Hou, Xiaoyi Liu, Lilin Chen
Abstract
Abstract To address the problem of object miss detection and object false detection in single threshold-Non-Maximum Suppression algorithm, this paper proposed a GDT-NMS (Generalized Intersection over Union Dual Threshold NMS, GDT-NMS)) algorithm which using GIoU(Generalized Intersection over Union). Using the GIoU indicator computing the similarity between objects, can better describe the relative position and overlap between the objects. And we proposed the dual-threshold NMS algorithm, which can balance the relationship between the object missed detection problem and the object false detection problem, reduce “false positive example” problem. By nonlinearly processing the weight function, the object is better distinguished. The algorithm uses Faster R-CNN as the detector. The experimental results show that the improved algorithm has outstanding performance.