soft nms
在目标检测中,非极大值抑制是重要的组成部分,然而它存在一个问题:他会把相邻的边界框iou大于一定阈值的直接置成0,所以可能会导致丢框现象出现,会降低评测指标。论文提出的方法,是会降低和M iou大于一定阈值的score,还是会在rank list里,并不是直接删除
A major issue with non-maxinmum suppresion is that it sets the score for neighboring detections to zero.Thus,if an object was actually presnet in that overlap threshold,it would be missed and this would lead to a drop in average precision.However,if we lower the detection scores as a function of its overlap with M,it would still be in the ranked list,although with a lower confidence.We show an illustration of the problem in Fig1.
原始nms公式
soft nms公式
这个公式是不连续的,所以提出了下面这个公式
实验结果
论文里,设置Nt是0.3,
σ
\sigma
σ设置成0.5.在coco数据集上,R-FCN和Faster RCNN分别提升了1.3%和1%。在Pascal voc 07数据集上,Faster RCNN和RFCN都提升了1.7%,但是对于SSD和Yolo v2只提升了0.5,这是因为sofnms建立在高召回的基础上
softnms对与重叠物体具有很好的适应性,就像fig 1所示那样,论文也提到对于斑马,大象等性能提升很大,而对于烤面包机,吹风机提升很小,因为他们不太可能在同一图像中出现多次