The test results only can be reproduced with the true test mask (not consistent with the paper)

[zysxmu]

Hi,
I have used the pre-trained models, but I found that the paper's test performances only can be reproduced with the true test mask.
That is, use these true mask in 'ISTD_Dataset/test/test_B/'.

However, if use the detected mask: test_mask_bdrar.zip provided by this repo. The results are:

While the paper says the results are obtained with the detected mask instead of real mask.