why use target/label to reconstruct image when test?

In inference time , why use the label to reconstruct the image while not use the prediction of the model and the digit capsules as the input of reconstruction module?
If the predict is wrong, the reconstruction module will reconstruct the wrong class of image?