NMS_非极大值抑制的作用

参考链接：NMS(非极大值抑制)

NMS: non maximum suppression

翻译为“非极大值抑制”，为什么不翻译成最大值抑制呢？maximum可以翻译为“最大值”，也可以翻译成“极大值”，所以翻译成极大值或者最大值一定要看这个值的含义。

极大值和最大值的区别就是，极大值是局部最大值。

NMS的作用：去掉detection任务重复的候选框，只留下预测概率值最大的候选框最为最终预测的结果（非极大值抑制）。

实现代码如下，来自链接：

 1 # 非极大值抑制
 2 def nms(bboxes, scores, score_thresh, nms_thresh, pre_nms_topk, i=0, c=0):
 3     """
 4     nms
 5     """
 6     inds = np.argsort(scores)
 7     inds = inds[::-1]
 8     keep_inds = []
 9     while(len(inds) > 0):
10         cur_ind = inds[0]
11         cur_score = scores[cur_ind]
12         # if score of the box is less than score_thresh, just drop it
13         if cur_score < score_thresh:
14             break
15 
16         keep = True
17         for ind in keep_inds:
18             current_box = bboxes[cur_ind]
19             remain_box = bboxes[ind]
20             iou = box_iou_xyxy(current_box, remain_box)
21             if iou > nms_thresh:
22                 keep = False
23                 break
24         if i == 0 and c == 4 and cur_ind == 951:
25             print('suppressed, ', keep, i, c, cur_ind, ind, iou)
26         if keep:
27             keep_inds.append(cur_ind)
28         inds = inds[1:]
29 
30     return np.array(keep_inds)
31 
32 # 多分类非极大值抑制
33 def multiclass_nms(bboxes, scores, score_thresh=0.01, nms_thresh=0.45, pre_nms_topk=1000, pos_nms_topk=100):
34     """
35     This is for multiclass_nms
36     """
37     batch_size = bboxes.shape[0]
38     class_num = scores.shape[1]
39     rets = []
40     for i in range(batch_size):
41         bboxes_i = bboxes[i]
42         scores_i = scores[i]
43         ret = []
44         for c in range(class_num):
45             scores_i_c = scores_i[c]
46             keep_inds = nms(bboxes_i, scores_i_c, score_thresh, nms_thresh, pre_nms_topk, i=i, c=c)
47             if len(keep_inds) < 1:
48                 continue
49             keep_bboxes = bboxes_i[keep_inds]
50             keep_scores = scores_i_c[keep_inds]
51             keep_results = np.zeros([keep_scores.shape[0], 6])
52             keep_results[:, 0] = c
53             keep_results[:, 1] = keep_scores[:]
54             keep_results[:, 2:6] = keep_bboxes[:, :]
55             ret.append(keep_results)
56         if len(ret) < 1:
57             rets.append(ret)
58             continue
59         ret_i = np.concatenate(ret, axis=0)
60         scores_i = ret_i[:, 1]
61         if len(scores_i) > pos_nms_topk:
62             inds = np.argsort(scores_i)[::-1]
63             inds = inds[:pos_nms_topk]
64             ret_i = ret_i[inds]
65 
66         rets.append(ret_i)
67 
68     return rets