0
$\begingroup$

In my experience it is quite commonplace to alter the loss function used when training a neural network for segmentation to ignore the contribution to the loss of unlabelled pixels. There are a few ways to implement it, but the one I have used prior is to alter network predictions to always predict truly unlabelled pixels correctly. In psuedo-code

predicted[where(labels == unlabelled)] = unlabelled

Is there a common, or most common, term for this?

I am attempting to add the technique itself as part of a literature review, though I can't find any examples which "directly" reference the technique. I can really only find articles that use the technique, but really make no mention of it.

$\endgroup$
0
$\begingroup$

You could do almost anything to implement this -- you could set the predictions to 0, or 1, or any value for the unlabelled predictions, and the training would still work, because the autodiff engine of whatever neural network library you're using would see that those predictions are not a function of your paramters $\theta$, and therefore the update wouldn't depend on them.

In particular, setting the predicted values to the ground truth is convenient, because the resulting loss is "average cross-entropy (or other loss) of the labelled pixels", rather than "average cross-entropy of labelled pixels, plus some arbitrary constant".

A common term for this is "masking". The loss of those pixels is "masked out", because they are set to 0 / ignored.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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