NRS: Non-Reliability Suppression for False Positives Elimination in YOLOv3

Ren Wang, Hyuk-Jae Lee, Jin-Sung Kim

Abstract: Non-maximum suppression (NMS) is essential as a post-processing step to eliminate the false positives caused by redundant detections for a single object. However, NMS cannot eliminate all the redundant detections which have different classification results. In order to address this problem, this paper proposes a novel post-processing step, non-reliability suppression (NRS). On PASCAL VOC, the proposed NRS reduces 28.4% false positives with only a 0.3% decrease of true positives for a YOLOv3 detector when compared with the traditional greedy-NMS. Because the computational complexity of NRS is low, the addition of NRS does not require much computational burden.