Prediction to image

[lfangyu09]

Hi,

Thanks a lot for providing this very useful code. I have a question about your code. Because I saw you used different thresholds to binarize the predicted image, I wonder how the prediction is changed to an image? prediction of the model -> sigmoid -> *255 to array -> image? Thanks!

[yhlleo]

Hi, @lfangyu09

Do you mean that how we get the predicted images?

If yes, we call the function model.test() first, which indeed run forward(). Then you can find the predicted results from the model: codes. Given the outputs of sigmoid function are with the range of [0, 1], I transformed it to be with the range of [-1, 1] (util. tensor2im requires the input images are with the range [-1, 1]).

In this inference procedure, we don't binarize the image with any thresholds. The thresholds are selected through the PR curve in the evaluation procedure.

[lfangyu09]

Hi,

Thanks a lot for your reply. If you change the outputs into the range of [-1,1], how do you save these outputs into images? Because I want to evaluate the output by using precision, recall, and so on. Thanks!

[yhlleo]

util.tensor2im converts the image with the range of [-1, 1] to [0, 255].

Then, you can evaluate the precision and recall with the saved images.