The paper introduces Soft-NMS (Non-maximum Suppression), a novel algorithm designed to improve object detection performance by refining the non-maximum suppression (NMS) step. Traditional NMS sets the scores of overlapping detection boxes to zero, which can lead to missed detections if the overlap is within a predefined threshold. Soft-NMS, however, decays the scores of these overlapping boxes as a continuous function of their overlap, preventing the elimination of any object. This approach maintains the detection scores while reducing their confidence, thus avoiding false negatives. The authors demonstrate that Soft-NMS consistently improves the mAP metric on standard datasets like PASCAL VOC 2007 and MS-COCO, achieving gains of 1.7% and 1.3% respectively for R-FCN and Faster-RCNN. The computational complexity of Soft-NMS is the same as traditional NMS, making it efficient and easy to integrate into existing object detection pipelines. The method is publicly available on GitHub.The paper introduces Soft-NMS (Non-maximum Suppression), a novel algorithm designed to improve object detection performance by refining the non-maximum suppression (NMS) step. Traditional NMS sets the scores of overlapping detection boxes to zero, which can lead to missed detections if the overlap is within a predefined threshold. Soft-NMS, however, decays the scores of these overlapping boxes as a continuous function of their overlap, preventing the elimination of any object. This approach maintains the detection scores while reducing their confidence, thus avoiding false negatives. The authors demonstrate that Soft-NMS consistently improves the mAP metric on standard datasets like PASCAL VOC 2007 and MS-COCO, achieving gains of 1.7% and 1.3% respectively for R-FCN and Faster-RCNN. The computational complexity of Soft-NMS is the same as traditional NMS, making it efficient and easy to integrate into existing object detection pipelines. The method is publicly available on GitHub.