nms 算法演示(附代码)
NMS_DEMO 动图演示
”’一共9个框,前4列是坐标(xyxy格式),*后一列是置信度”’
boxes=np.array([
[100,100,210,210,0.72],#0
[280,290,420,420,0.8],#1
[220,220,320,330,0.92],#2
[105,90,220,210,0.71],#3
[230,240,325,330,0.81],#4
[305,300,420,420,0.9],#5
[215,225,305,328,0.6],#6
[150,260,290,400,0.99],#7
[102,108,208,208,0.72]])#8
”’按置信度从高到低排序(框对应的索引)”’
order: [7 2 5 4 1 8 0 3 6]
左图是原始的框(没经过nms操作的框):
黄色填充的框是按置信度排序后,被选中保留的框
右图是nms操作后的框:
每选中保留一个框,就会计算它跟余下的框的iou,iou大于0.7(人为设置)则剔除
注意:这里余下的框是不包括之前已经确定保留的框
左图的题目keep表示保留的框的id,score表示保留的框的置信度
右图的题目keep表示保留的框的id,delete表示需要剔除的框(保留的框与其iou>0.7)
介绍一下步骤:
所有框按置信度从高到低进行排序:[7 2 5 4 1 8 0 3 6] (框的索引)
从置信度*高的那个框(7)开始,(7)已经保留,那么就会跟(0)~(6)、(8)剩下的8个框比较iou,iou都<0.7,全部保留
接着确定置信度排第2的框(2),(7,2)已经保留,那么就会跟(0)、(1)、(3)、(4)~(6)、(8)剩下的7个框比较iou,其中(4)、(6)iou>0.7,剔除
接着确定置信度排第3的框(5),(7,2,5)已经保留,那么就会跟(0)、(1)、(3)、(8)剩下的4个框比较iou,其中(1)iou>0.7,剔除
接着确定置信度排第4的框(8),(7,2,5,8)已经保留,那么就会跟(0)、(3)剩下的2个框比较iou,其中(0)、(3)iou>0.7,剔除
完整代码
import numpy as np
import matplotlib.pyplot as plt
def py_cpu_nms(nms_show, dets, thresh):# nms操作
“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]
areas = (y2-y1+1) * (x2-x1+1) #每一个检测框的面积
print(“all_areas:”,areas)
order = scores.argsort()[::-1]# 按照score置信度降序排序
print(“order:”,order)
keep = [] # 保留的结果框集合
k=0
show_boxid = order
print(“show_boxid”,show_boxid)
while order.size > 0:
i = order[0] # every time the first is the biggst, and add it directly
keep.append(i) # 保留该类剩余box中得分*高的一个
print(“\nkeep(被留下来的框id):”,keep)
# 得到相交区域,比左上大和比右下小
”’np.maximum(X, Y, out=None) X和Y逐位进行比较,选择*大值”’
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:]])
print(“xx1,yy1:”,xx1,yy1)
print(“xx2,yy2:”,xx2,yy2)
print(“w,h:”,xx2-xx1+1,yy2-yy1+1)
#plt.scatter([xx1,xx2], [yy1,yy2], s=50, c=’y’)#框左上角蓝色
# 计算相交的面积,不重叠时面积为0
w = np.maximum(0, xx2-xx1+1) # the weights of overlap
h = np.maximum(0, yy2-yy1+1) # the height of overlap
inter = w*h
print(“w*h:”,inter)
# 计算IoU:重叠面积 /(面积1+面积2-重叠面积)
iou = inter / (areas[i]+areas[order[1:]] – inter)
print(“—iou:”,iou)
# 保留IoU小于阈值的box
”’去掉keep剩下的框按顺序重新排序”’
indx = np.where(iou<=thresh)[0]
dedx = np.where(iou>thresh)[0]
print(“order[0]:”,order[0])
show_boxid = np.delete(show_boxid,np.where(show_boxid==order[0])[0],axis = 0)
print(“show_boxid:”,np.append(keep,show_boxid))
delete_boxid = show_boxid[dedx]
show_boxid = show_boxid[indx]
#print(“indx:”,indx)
print(“after_iou_show_boxid:”,np.append(keep,show_boxid))
print(“after_iou_delete_boxid:”,delete_boxid,’\n’)
k=k+1
”’绘制动图”’
ax1 = nms_show.add_subplot(1,2,1)
ax1.set_title(‘begin_nms {} \nkeep:{} score:{}({})’.format(k,order[0],scores[order[0]],k))
plot_bbox(dets, ‘k’, show_ids=np.arange(9) , keep_id = order[0])
ax2 = nms_show.add_subplot(1,2,2)
ax2.set_title(‘after_nms {} \nkeep:{} delete:{}(iou={})’.format(k,order[0],delete_boxid,iou[dedx]))
plot_bbox(dets[np.append(show_boxid,keep)], ‘b’, np.append(show_boxid,keep))
plt.pause(5)
ax1.remove()
ax2.remove()
”’置信度排前的数值给取出,剩下的数构成新的数组”’
order = order[indx+1]
print(“———–afer_order:”,order,’———–‘)
return keep
def plot_bbox(dets, c=’k’, show_ids=[],keep_id=0):
x1 = dets[:,0]
y1 = dets[:,1]
x2 = dets[:,2]
y2 = dets[:,3]
score=dets[:,4]
#print(dets.shape)
#plt.scatter(x1, y1, s=25, c=’b’, alpha=0.6)#框左上角蓝色
#plt.scatter(x2, y2, s=25, c=’r’, alpha=0.6)#框右下角红色
plt.plot([x1,x2], [y1,y1], c)
plt.plot([x1,x1], [y1,y2], c)
plt.plot([x1,x2], [y2,y2], c)
plt.plot([x2,x2], [y1,y2], c)
plt.xlim((60,450))
plt.ylim((450,60))
”’改变坐标轴位置”’
ax = plt.gca()
ax.spines[“top”].set_color(“k”)
ax.xaxis.set_ticks_position(“top”)
for i in range(len(show_ids)):
plt.text(x1[i], y1[i]+7, “(%d)%.2f”%(show_ids[i],score[i]), \
fontdict={‘size’: 10, ‘color’: ‘r’},bbox={‘facecolor’:’blue’, ‘alpha’:0.1})
if keep_id != 0:
ax.add_patch(plt.Rectangle((x1[keep_id], y1[keep_id]), x2[keep_id]-x1[keep_id]+1, y2[keep_id]-y1[keep_id]+1,
color=”y”, fill=True, linewidth=2))
def main():
boxes=np.array([
[100,100,210,210,0.72],#0
[280,290,420,420,0.8],#1
[220,220,320,330,0.92],#2
[105,90,220,210,0.71],#3
[230,240,325,330,0.81],#4
[305,300,420,420,0.9],#5
[215,225,305,328,0.6],#6
[150,260,290,400,0.99],#7
[102,108,208,208,0.72]])#8 #9个框
plt.ion()
fig = plt.figure(figsize=[14,9])
ax1 = plt.subplot(1,2,1)
ax1.set_title(‘before_nms’)
ax2 = plt.subplot(1,2,2)
ax2.set_title(‘after_nms’)
plt.sca(ax1)# 选择子图1
plot_bbox(boxes,’k’,show_ids=np.arange(9),keep_id=0) # before nms
keep = py_cpu_nms(fig, boxes, thresh=0.7)
print(“last_keep:”,keep)
plt.ioff()
plt.pause(2)
plt.close(‘all’)
if __name__ ==”__main__”:
main()