prediction confidence score?

[xuebinqin]

prediction confidence score?

[xuebinqin]

Thanks for your interests. The prediction confidence is exactly what you mentioned in the following question (1). But we didn't explicitly compute the prediction confidence.
(1) Yes, it does. When you compute the difference between the prediction, you will get a error map, which was minimized by the training process.
(2) It sounds like a GAN: use a network to get prediction and use another one to evaluate the the prediction quality. But for segmentation tasks, the losses should be calculated pixel-wise not image wise (keep the error map not the mean error of the error map) because the predictions of some pixels in an image may be good others are not. I guess you want to recognized these bad predicted pixels not images.