I am trying to use unet to learn pixel-wise regression from one image to one groundtruth with the same image size. The network seems to focus too much on the edge of the image, and it does not learn much on the inside. Loss is only calculated inside the valid region. The loss is in a good looking decreasing shape. The examples of groundtruth and prediction are shown below. Does anyone have any recommendation on the issue?


Groundtruth image

enter image description here

Prediction image

enter image description here Epoch loss

  • $\begingroup$ Could you show the input image? Also share the architecture? $\endgroup$ Sep 14, 2023 at 6:45
  • $\begingroup$ show the images as rgb/b&w not as 3d model $\endgroup$
    – Alberto
    Sep 14, 2023 at 10:55
  • $\begingroup$ What loss function are you using? $\endgroup$ Sep 14, 2023 at 15:37


You must log in to answer this question.