This paper proposes Soft-NMS, a modified version of the traditional non-maximum suppression (NMS) algorithm for object detection. Unlike traditional NMS, which suppresses detection boxes with significant overlap with the highest-scoring box, Soft-NMS decays the scores of overlapping boxes as a continuous function of their overlap with the highest-scoring box. This approach ensures that no object is eliminated, and it leads to consistent improvements in the COCO-style mAP metric on standard datasets like PASCAL VOC 2007 and MS-COCO. Soft-NMS achieves improvements of 1.7% for both R-FCN and Faster-RCNN on PASCAL VOC 2007, and 1.3% and 1.1% for R-FCN and Faster-RCNN on MS-COCO. It also improves state-of-the-art performance in object detection from 39.8% to 40.9% with a single model. Soft-NMS has the same computational complexity as traditional NMS and can be easily integrated into any object detection pipeline without requiring additional training or hyper-parameters. The algorithm is implemented using a Gaussian penalty function, which continuously penalizes overlapping boxes based on their overlap with the highest-scoring box. Soft-NMS is shown to perform better than traditional NMS in various scenarios, including when there is a significant overlap between detection boxes and when the overlap threshold is increased. The algorithm is also shown to improve precision and recall in object detection, particularly for objects that are less likely to co-occur in the same image. The code for Soft-NMS is publicly available on GitHub.This paper proposes Soft-NMS, a modified version of the traditional non-maximum suppression (NMS) algorithm for object detection. Unlike traditional NMS, which suppresses detection boxes with significant overlap with the highest-scoring box, Soft-NMS decays the scores of overlapping boxes as a continuous function of their overlap with the highest-scoring box. This approach ensures that no object is eliminated, and it leads to consistent improvements in the COCO-style mAP metric on standard datasets like PASCAL VOC 2007 and MS-COCO. Soft-NMS achieves improvements of 1.7% for both R-FCN and Faster-RCNN on PASCAL VOC 2007, and 1.3% and 1.1% for R-FCN and Faster-RCNN on MS-COCO. It also improves state-of-the-art performance in object detection from 39.8% to 40.9% with a single model. Soft-NMS has the same computational complexity as traditional NMS and can be easily integrated into any object detection pipeline without requiring additional training or hyper-parameters. The algorithm is implemented using a Gaussian penalty function, which continuously penalizes overlapping boxes based on their overlap with the highest-scoring box. Soft-NMS is shown to perform better than traditional NMS in various scenarios, including when there is a significant overlap between detection boxes and when the overlap threshold is increased. The algorithm is also shown to improve precision and recall in object detection, particularly for objects that are less likely to co-occur in the same image. The code for Soft-NMS is publicly available on GitHub.