在Keras中实施IOU措施

Question

我实现了如下IOU：

import numpy as np

EPS = 1e-12

def get_iou( gt , pr , n_classes ):
    class_wise = np.zeros(n_classes)
    for cl in range(n_classes):
        intersection = np.sum(( gt == cl )*( pr == cl ))
        union = np.sum(np.maximum( ( gt == cl ) , ( pr == cl ) ))
        iou = float(intersection)/( union + EPS )
        class_wise[ cl ] = iou
    return class_wise

然后像这样使用它：

for l,y in zip(gt_label_list,network_prediction):
    gt=cv2.imread(l,0)
    pr=y
    iou.append(get_iou(gt,pr,num_classes=20))

ious = np.array(iou)
print("Class wise IoU near"  ,  np.mean(ious , axis=0 ))
print("Total  IoU near"  ,  np.mean(ious ))

但是，即使我为相同的图像（cityscapes 500验证图像集）馈送gt和pr时，总IOU也不合理：

Class wise IoU near [0.968 0.932 0.982 0.404 0.378 0.982 0.582 0.946 0.972 0.482 0.886 0.804 0.506 0.958 0.16  0.15  0.044 0.186 0.69  1.   ]

Total  IoU near 0.6505999999999998

我知道有些类并没有出现在一些图片中，但这篇文章是如何报告MIoUs超过70%的呢