Can you give a boundary loss for binary segmentation, thanks

[ustczhouyu]

Can you give a boundary loss for binary segmentation, or can you tell me how to modify the code to use it for binary segmentation， thanks

[yiskw713]

Hi, ustczhouyu

Sorry for the delay in replying.
How about converting binary prediction to two-class prediction?

import torch
n_batch = 4
H = 112
W = 112

# prediction for binary cross entropy
binary_pred = torch.rand(n_batch, 1, H, W)

# background prediction
bg_pred = torch.ones_like(binary) - binary_pred

# foreground prediction
fg_pred = binary_pred.clone()

# 2 class prediction. shape => (n_batch, 2, H, W)
pred = torch.cat([bg_pred, fg_pred], dim=1)

[NguyenHungHust256]

Hi, ustczhouyu

Sorry for the delay in replying.
How about converting binary prediction to two-class prediction?

import torch
n_batch = 4
H = 112
W = 112

# prediction for binary cross entropy
binary_pred = torch.rand(n_batch, 1, H, W)

# background prediction
bg_pred = torch.ones_like(binary) - binary_pred

# foreground prediction
fg_pred = binary_pred.clone()

# 2 class prediction. shape => (n_batch, 2, H, W)
pred = torch.cat([bg_pred, fg_pred], dim=1)

Can you explain code. I don't know why you implement this code

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713

[ustczhouyu]

Thanks for sharing your code. When doing binary segmentation (one class is the object, the other class is the background), I want to know this function: loss = criterion(y['out'], gt), where 'gt' should be the ground-truth of the whole instance or only ground-truth of the boundary? @yiskw713