1
$\begingroup$

I have a question regarding the convolutional neural network known as U-Net (see link below) and hope somebody can help me out.

https://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/

In the paper, the authors mention a map with a pixel-wise loss weight to force the network to learn the border pixels. I assume this weight map is intended to initialize the weights but I have failed to find the code showing how these weight maps are used to initialize the weights of the U-Net. I am not familiar with Caffe and instead use Keras to implement the U-Net. My current code is partially based on https://github.com/zhixuhao/unet. In this implementation, no pre-computed weight map is included

Can somebody please point out to me how this weight map described in the paper was used to initialize the weights of the U-Net?

Any Help would be greatly appreciated!

$\endgroup$
1
$\begingroup$

The weight map is not related to the network weights but to the sample weights: during the training, you feed three pieces of information to the network, 1) the pixel intensities, 2) the pixel labels, 3) the pixel weights.

The loss function used is simply weighted cross entropy.

$\endgroup$
  • $\begingroup$ Thanks for your reply. But how are these pixel weights passed into the network? I was not able to find out how to pass in pixel weights into a Keras implemented ConvNet? $\endgroup$ – disputator1991 Mar 20 '18 at 14:09
  • $\begingroup$ Probably along with the labels blob. Also, I don't know if Keras implements a weighted cross entropy directly, maybe you will have to implement it yourself like this $\endgroup$ – Jan Kukacka Mar 20 '18 at 15:12

Not the answer you're looking for? Browse other questions tagged or ask your own question.