About BCEloss with abnormal score

[hanwenxu1]

I have know that we can calculate the abnormal scores by the number of abnormal blocks. But I notice that a sigmoid layer and binary cross-entropy been used in the model. Is there any dismatch between the linear score and a binary loss?

Thanks for your reply!

[ziyye]

The paper states that "Images assigned with abnormal labels are positive samples while normal ones are negative samples.", and the score is in range of [0, 1]. So, it's a binary classification task in fact. But the author said that the number of abnormal bounding boxes as the score, I feel confused ...