2020-07-14 目标检测

1. NMS算法：

• 代码实现

# Bounding Boxes: boxes [B1,B2,...,Bn]
# B = [x1,y1,x2,y2]
# (x1,y1)         
#          (x2,y2)
# Confidences: scores [S1,S2,...,Sn]
# NMS Threshold: threshold 
# Outpus Boxes:keep []
def NMS(boxes, scores, threshold)
    proposal = sorted(list(zip(boxes, scores)), key = lambda x:x[1], reverse = True)
    keep = [ ]
    while proposal:
        max_proposal = proposal[0]
        keep.append(proposal[0])
        proposal = proposal[1:]
        for index, pro in enumerate(proposal):
            if IOU(max_proposal, pro) >= threshold:
                 proposal.pop(index)
    return keep

def IOU(pro1, pro2):
    box1 = pro1[0]
    box2 = pro2[0]
    area1 = (box1[2]-box1[0])*(box1[3]-box1[1])
    area2 = (box2[2]-box2[0])*(box2[3]-box2[1])
    xx1 = max(box1[0], box2[0])
    yy1 = max(box1[1], box2[1]) # 找出更大的x1,y1
    xx2 = min(box1[2], box2[2])
    yy2 = min(box1[3], box3[3]) # 找出更小的x2,y2
    w = max(0, xx2-xx1+1)
    h = max(0, yy2-yy1+1)
    inter = w*h
    over = inter / (area1 + area2 - inter)
    return over

• 可能存在的问题：NMS的阈值不容易确定，阈值太小，当两个物体的框重叠较多，可能会出现漏检；阈值太大，容易增大误检。
• 解决办法：Soft-NMS, 不直接删除所有IoU大于阈值的框，而是降低其置信度。

2. mAP:

P：precision，预测正确的个数/测试总个数
AP：average precision，每一类别P值的平均值
MAP：mean average precision，对所有类别的AP取均值

(1) mAP的计算

mAP.png

(2) MS COCO mAP

MS COCO mAP.png