【AI达人创造营第二期】目标检测中的loss函数

# 三种loss的简单测试import paddleimport numpy as npimport paddle.nn.functional as F

x = np.random.rand(10,10)
y = np.random.rand(10,10)

x = paddle.to_tensor(x)
y = paddle.to_tensor(y)

L1loss = F.l1_loss(x,y)
L2loss = F.mse_loss(x,y)
SmoothL1loss = F.smooth_l1_loss(x,y)print('L1loss:',L1loss)print('L2loss:',L2loss)print('Smoothloss:',SmoothL1loss)

import cv2import math 
import matplotlib.pyplot as plt 
import numpy as npimport paddle 
def draw_rect(img,box,color):   # 画box
    pt1 = (int(box[0]), int(box[1]))
    pt2 = (int(box[0] + box[2]), int(box[1] + box[3]))
    img = cv2.rectangle(img, pt1, pt2, color, thickness=3)    return imgdef show(img_path,bbox):
    cbox = [min(bbox[0],gt[0]),                  
        min(bbox[1],gt[1]),        max(bbox[0]+bbox[2],gt[0]+gt[2])-min(bbox[0],gt[0]),        max(bbox[1]+bbox[3],gt[1]+gt[3])-min(bbox[1],gt[1])]
    plt.title(img_path)
    img = cv2.imread(img_path)
    img = draw_rect(img,gt,(0,0,255))
    img = draw_rect(img,bbox,(0,255,0))
    img1 = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
    plt.imshow(img1)
    plt.show()# ltwh转成xxyydef convertltwh2xxyy(box): 
    x1= box[0]
    x2= box[0] + box[2]
    y1= box[1]
    y2= box[1] + box[3]    return x1,x2,y1,y2def computeIoU(box1,box2, GIoU=False, DIoU=False, CIoU=False, xxyy=False): # box1:gt,box2:预测
    if xxyy:
        b1x1,b1x2,b1y1,b1y2=box1[0],box1[1],box1[2],box1[3]
        b2x1,b2x2,b2y1,b2y2=box2[0],box2[1],box2[2],box2[3]    else: 
        b1x1,b1x2,b1y1,b1y2 = convertltwh2xxyy(box1)
        b2x1,b2x2,b2y1,b2y2 = convertltwh2xxyy(box2)

    inter = (min(b1x2,b2x2)-max(b1x1,b2x1)) * (min(b1y2,b2y2)-max(b1y1,b2y1))  #求交集
    if inter < 0 : 
        inter = 0
    w1,h2 = b1x2-b1x1,b1y2-b1y1
    w2,h2 = b2x2-b2x1,b2y2-b2y1
    union = w1*h2 + w2*h2 - inter   # 求并集
    iou = inter/union    if CIoU or DIoU or GIoU:
        w_c = max(b1x2,b2x2)-min(b1x1,b2x1)  #最小矩形的长宽
        h_c = max(b1y2,b2y2)-min(b1y1,b2y1) 
        if CIoU or DIoU:
           c_2 = w_c**2 +h_c**2  # 对角线的平方
           rho2 =  (((b1x1+b1x2)/2)-((b2x2+b2x1)/2))**2+(((b1y1+b1y2)/2)-((b2y2+b2y1)/2))**2 #中心点距离
           if CIoU:
               v = (4/math.pi**2)*(math.atan(w1/h2) - math.atan(w2/h2))**2   #长宽比信息
               with paddle.no_grad():
                   alpha = v / (1-iou+v)               return iou - rho2 / c_2 - v * alpha # CIoU
           return iou-rho2/c_2 #DIoU
        area_c = w_c * h_c        return iou - ((area_c - union)/area_c) #GIoU
    return iouif __name__ == '__main__':
    img_path = 'work/00000001.jpg'
    label_path = 'work/groundtruth.txt'
    gt =  np.loadtxt(label_path,delimiter=',')   #[l,t,w,h]
    bbox_list = [[1200,450,500,600],[350,440,400,300],[250,400,450,300],[400,500,200,150]]    for i in range(len(bbox_list)):   
        bbox = bbox_list[i]
        show(img_path,bbox)        print("IoU:",format(computeIoU(gt,bbox),'.4f'))

    for i in range(len(bbox_list)):   
        bbox = bbox_list[i]
        show(img_path,bbox)        print("GIoU:",format(computeIoU(gt,bbox,GIoU=True),'.6f'))

    for i in range(len(bbox_list)):   
        bbox = bbox_list[i]
        show(img_path,bbox)        print("DIoU:",format(computeIoU(gt,bbox,DIoU=True),'.6f'))        print("CIoU:",format(computeIoU(gt,bbox,CIoU=True),'.6f'))

# pred_iou是head输出的预测值def iou_aware(pred_iou,gt,box, GIoU=False, DIoU=False, CIoU=False):
    iou = computeIoU(gt,box, GIoU=False, DIoU=False, CIoU=False)
    iou=paddle.to_tensor(iou)
    pred_iou=paddle.to_tensor(pred_iou)    with paddle.no_grad():
        loss=F.binary_cross_entropy_with_logits(pred_iou,iou)    return loss 

for i in range(len(bbox_list)):
        p = 0.8   # 预测iou值 
        bbox = bbox_list[i]
        show(img_path,bbox)        print("IoUawareloss:",iou_aware(p,gt,bbox))