Object Detection - An Overview of Object Detection

Target Detection

Commonly used open source datasets

PASCAL VOC dataset

download link

MS COCO dataset

Commonly used evaluation indicators

IOU (intersection over union)

import numpy as np

def IOU(box1,box2,wh=False):
    #判断表示方式
    if wh==False:
        #极坐标表示
        xmin1,ymin1,xmax1,ymax1=box1
        xmin2,ymin2,xmax2,ymax2=box2
    else:
        #中心点坐标表示
        #第一框
        xmin1,ymin1=int(box1[0]-box1[2]/2.0),int(box1[1]-box1[3]/2.0)
        xmax1,ymax1=int(box1[0]+box1[2]/2.0),int(box1[1]+box1[3]/2.0)
        #第二框
        xmin2,ymin2=int(box2[0]-box2[2]/2.0),int(box2[1]-box2[3]/2.0)
        xmax2,ymax2=int(box2[0]+box2[2]/2.0),int(box2[1]+box2[3]/2.0)
    #获取交集的左上角和右下角坐标
    xx1=np.max(xmin1,xmin2)
    yy1=np.max(ymin1,ymin2)
    xx2=np.min(xmax2,xmax1)
    yy2=np.min(ymax1,ymax2)
    #计算交集面积
    inter_area=(np.max(0,xx2-xx1))*(np.max(0,yy2-yy1))
    #计算并的面积
    area1=(xmax1-xmin1)*(ymax1-ymin1)
    area2=(xmax2-xmin2)*(ymax2-ymin2)
    union_area=area1+area2-inter_area
    #IOU
    IOU=inter_area/(union_area+1e-6)
    return IOU

import matplotlib.pyplot as plt
import matplotlib.patches as patches #给图像打上矩形框

#真是框 预测框
true_box=[100,35,398,400]
pre_box=[40,150,355,398]
#将框绘制在图像上
img=plt.imread('dog.jpeg')
fig=plt.imshow(img)
#将真实框和预测框绘制在图像上
fig.axes.add_patch(plt.Rectangle((true_box[0],
                                  true_box[1],
                                  width=true_box[2]-true_box[1],
                                  height=true_box[3]-true_box[1],
                                  fill=False,
                                  edgcolor='blue',
                                  linewidth=2)))
fig.axes.add_patch(plt.Rectangle((pre_box[0],
                                  pre_box[1],
                                  width=pre_box[2]-pre_box[1],
                                  height=pre_box[3]-pre_box[1],
                                  fill=False,
                                  edgcolor='red',
                                  linewidth=2)))

IOU(ture_box,pre_box)

0.5114435907762924

mAP（Mean Average Precision）

Before voc2010, calculate the presion corresponding to Recall11 points,
after calculating the average of voc2010, use the area under the line of the piecewise function

NMS (Non-Maximum Suppression)

Remove redundant detection frames and keep the best one

def nms(boxes,score,thre):
    #容错处理
    if (len(boxes)==0):
        return [],[]
    #类型转换
    #box使用极坐标
    boxes=np.array(boxes)
    score=np.array(score)
    #获取左上角和右下角坐标
    x1=boxes[:,0]
    y1=boxes[:,1]
    x2=boxes[:,2]
    y2=boxes[:,3]
    #计算面积
    areas=(x2-x1)*(y2-y1)
    #NMS
    picked_boxes=[]
    picked_socre=[]
    #排序:小->大
    order=np.argsort(score)
    while order.size>0:
        #获取score最大的索引
        index=order[-1]
        #保留下来
        picked_boxes.append(boxes[index])
        picked_score.append(score[index])
        #计算IOU
        #交的面积
        x11=np.maximum(x1[index],x1[order[:-1]])
        y11=np.maximum(y1[index],y1[order[:-1]])
        x22=np.maximum(x2[index],x2[order[:-1]])
        y22=np.maximum(y2[index],y2[order[:-1]])
        w=np.maximum(0.0,x22-x11)
        h=np.maximum(0.0,y22-y11)
        inter_area=w*h
        #交并比
        iou=inter_area/(areas[index]+areas[order[:-1]]-inter_area)
        #删除冗余框
        keep_boxes=np.where(iou<thre)
        #更新order
        order=order[keep_boxes]
    return picked_boxes,picked_score

bounding=[(187,82,337,317),(150,67,305,282),(246,121,368,304)]
confidence_score=[0.9,0.65,0.8]
threshold=0.5

box,score=nms(bounding,confidence_score,threshold)

Classification of object detection methods