非极大值抑制（NMS）

• NMS：常用于目标检测中，将候选框中与高分候选框重合度大于某一阈值的候选框去掉。

• 算法流程：

  1. 令filter_boxes为待筛选的候选框的集合，初始化为所有候选框。令chosen_boxes为已经入选的候选框集合，初始化为空。对filter_boxes根据预测概率从高到底进行排序。
  2. 遍历filter_boxes
     1. 将filter_boxes中概率最高的filter_boxes[0]放入chosen_boxes。
     2. 计算filter_boxes[0]与filter_boxes剩余候选框的IoU，将重合度大于阈值的框去除。为什么只计算与filter_boxes[0]的IoU呢，因此与其他已经进入chosen_boxes的候选框已经进行过筛选了。
     3. 调回1直到filter_boxes为空。

• code（python）

import numpy as np

# boxes contain the coordinates and scores (x1, y1, x2, y2, score) of the boxes
def NMS(boxes, thres):
	x1 = boxes[:, 0]
	y1 = boxes[:, 1]
	x2 = boxes[:, 2]
	y2 = boxes[:, 3]
	scores = boxes[:, 4]
	
	areas = (x2-x1+1)*(y2-y1+1)
	filter_idxes = scores.argsort()[::-1] #sort the indexes according to the score in descending order

	chosen_idxes = [] #将chosen_boxes的下标初始化为空
	while filter_idxes.size() > 0:
		c_idx = filter_idxes[0]
		chosen_idxes.append(c_idx)
		filter_idxes = filter_idxes[1:]
		
		interaction_x1 = np.maximum(x1[c_idx], x1[filter_idxes])
		interaction_y1 = np.maximum(y1[c_idx], y1[filter_idxes])
		interaction_x2 = np.minimum(x2[c_idx], x2[filter_idxes])
		interaction_y2 = np.minimum(y2[c_idx], y2[filter_idxes])
		
		w = np.maximum(0., interaction_x2-interaction_x1+1) # 当两个候选框没有交集的时候，interaction_x2-interaction_x1<0
		h =  np.maximum(0., interaction_y2-interaction_y1+1)
	
		interaction = h * w
		union = area[c_idx] + area[filter_idxes] - interaction
		IoU = interaction / union
		
		remain_idx = np.where(IoU <= thres)[0] #选出在filter_idxes中IoU小于阈值的下标，这个下标是filter_idxes的下标！
		filter_idxes = filter_idxes[remain_idx]
	
	return chosen_idxes