转：非极大值抑制（Non-Maximum Suppression，NMS）

非极大值抑制（Non-Maximum Suppression，NMS）

单类别NMS的numpy实现

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

def py_cpu_nms(dets, thresh):     """Pure Python NMS baseline."""     #x1、y1、x2、y2、以及score赋值     x1 = dets[:, 0]     y1 = dets[:, 1]     x2 = dets[:, 2]     y2 = dets[:, 3]     scores = dets[:, 4]       #每一个检测框的面积     areas = (x2 - x1 + 1) * (y2 - y1 + 1)     #按照score置信度降序排序     order = scores.argsort()[::-1]       kept_bboxes = [] #保留的结果框集合     while order.size > 0:         i = order[0]         kept_bboxes.append(i) #保留该类剩余box中得分最高的一个         #得到相交区域,左上及右下         xx1 = np.maximum(x1[i], x1[order[1:]])         yy1 = np.maximum(y1[i], y1[order[1:]])         xx2 = np.minimum(x2[i], x2[order[1:]])         yy2 = np.minimum(y2[i], y2[order[1:]])           #计算相交的面积,不重叠时面积为0         w = np.maximum(0.0, xx2 - xx1 + 1)         h = np.maximum(0.0, yy2 - yy1 + 1)         inter = w * h         #计算IoU：重叠面积 /（面积1+面积2-重叠面积）         ovr = inter / (areas[i] + areas[order[1:]] - inter)         #保留IoU小于阈值的box         inds = np.where(ovr <= thresh)[0]         order = order[inds + 1] #因为ovr数组的长度比order数组少一个,所以这里要将所有下标后移一位           return kept_bboxes

　　

单类别NMS的pytorch实现

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

def _nms(self, bboxes, scores, threshold=0.5):     x1 = bboxes[:,0]     y1 = bboxes[:,1]     x2 = bboxes[:,2]     y2 = bboxes[:,3]     areas = (x2-x1)*(y2-y1)   # [N,] 每个bbox的面积     _, order = scores.sort(0, descending=True)    # 降序排列       kept_bboxes = []     while order.numel() > 0:       # torch.numel()返回张量元素个数         if order.numel() == 1:     # 保留框只剩一个             i = order.item()             kept_bboxes.append(i)             break         else:             i = order[0].item()    # 保留scores最大的那个框box[i]             kept_bboxes.append(i)           # 计算box[i]与其余各框的IOU(思路很好)         xx1 = x1[order[1:]].clamp(min=x1[i])   # [N-1,]         yy1 = y1[order[1:]].clamp(min=y1[i])         xx2 = x2[order[1:]].clamp(max=x2[i])         yy2 = y2[order[1:]].clamp(max=y2[i])         inter = (xx2-xx1).clamp(min=0) * (yy2-yy1).clamp(min=0)   # [N-1,]           iou = inter / (areas[i]+areas[order[1:]]-inter)  # [N-1,]         idx = (iou <= threshold).nonzero().squeeze() # 注意此时idx为[N-1,] 而order为[N,]         if idx.numel() == 0:             break         order = order[idx+1]  # 修补索引之间的差值     return torch.LongTensor(kept_bboxes)   # Pytorch的索引值为LongTensor