IoU Loss

突出前景

def _iou(pred, target):

1
2
3
4
5
6
7
8
9
10
b = pred.shape[0]
IoU = 0.0
for i in range(0,b):
    Iand1 = torch.sum(target[i,:,:,:]*pred[i,:,:,:])
    Ior1 = torch.sum(target[i,:,:,:]) + torch.sum(pred[i,:,:,:])-Iand1
    IoU1 = Iand1/Ior1

    IoU = IoU + (1-IoU1)    #因为要算的是错误的大小，所以要1-IoU

return IoU/b

EGNet中的边缘loss

Paper:http://mftp.mmcheng.net/Papers/19ICCV_EGNetSOD.pdf
GitHub:https://github.com/JXingZhao/EGNet/
在这里插入图片描述
在EGNet中，是把阈值为0.5的二值化label扔进边缘损失函数中
计算正常的loss部分是用0为阈值的二值化label

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
def load_edge_label(im):
    """
    pixels > 0.5 -> 1
    Load label image as 1 x height x width integer array of label indices.
    The leading singleton dimension is required by the loss.
    """
    label = np.array(im, dtype=np.float32)
    if len(label.shape) == 3:
        label = label[:,:,0]
    # label = cv2.resize(label, im_sz, interpolation=cv2.INTER_NEAREST)
    label = label / 255.
    label[np.where(label > 0.5)] = 1.   #  0.5当做阈值
    label = label[np.newaxis, ...]
    return label

def EGnet_edg(d,labels_v):
    target=load_edge_label(labels_v)
    # assert(d.size() == target.size())
    pos = torch.eq(target, 1).float()
    neg = torch.eq(target, 0).float()
    # ing = ((torch.gt(target, 0) & torch.lt(target, 1))).float()

    num_pos = torch.sum(pos)
    num_neg = torch.sum(neg)
    num_total = num_pos + num_neg

    alpha = num_neg  / num_total
    beta = 1.1 * num_pos  / num_total
    # target pixel = 1 -> weight beta
    # target pixel = 0 -> weight 1-beta
    weights = alpha * pos + beta * neg

    return F.binary_cross_entropy_with_logits(d, target, weights, reduction=None)

论文中是把label二值化的部分加在dataset部分，我想直接通过后处理来实现，但是是一些奇奇怪怪的报错。。。
下面贴上原本的load_edge_label部分：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
def load_edge_label(pah):
    """
    pixels > 0.5 -> 1
    Load label image as 1 x height x width integer array of label indices.
    The leading singleton dimension is required by the loss.
    """
    if not os.path.exists(pah):
        print('File Not Exists')
    im = Image.open(pah)
    label = np.array(im, dtype=np.float32)
    if len(label.shape) == 3:
        label = label[:,:,0]
    # label = cv2.resize(label, im_sz, interpolation=cv2.INTER_NEAREST)
    label = label / 255.
    label[np.where(label > 0.5)] = 1.
    label = label[np.newaxis, ...]
    return label

然后在读取dataload的时候读取一下处理后的edge_label，扔到EGnet_edg里面去算就好了

该论文在算loss的时候对loss进行大小的调整
nAveGrad=10

1
sal_loss = (sum(sal_loss1) + sum(sal_loss2)) / (nAveGrad * self.config.batch_size)