Design question

[urimerhav]

In your paper you describe InverseForm is used to compare b_pred with b_gt. The latter is obtained from running a Sobel filter on the ground truth segmentation . Is there any particular reason we don't do the same to obtainb_pred? Meaning, we could have just taken the predicted segmentation map, ran a sobel operation on that, and fed that as b_pred.

Put formally in your notation:

b_gt = sobel(y_gt)
So why not also
b_pred = sobel(y_pred)

[sborse3]

Hi @urimerhav, thanks for your question! We could enforce constraints on the semseg outputs by converting them to b_pred. However, this is slightly more difficult to converge; we have experimented with this possibility, finding underwhelming results. Feel free to perform these experiments on your end and let us know what you observe

[urimerhav]

Thanks a lot for your quick reply! That you got a negative result is exactly what I was curious to learn! That's great to know.

Do you have any insight on why this fails? Or it's just an empirical observation that's currenty uexplained.

[sborse3]

I feel a larger parameter search might help with better results. As we are adding a constraint to our generated segmentation map, we might observe gains but in a smaller parameter space, which would technically be hard to find. Till then, it's an empirical observation with the limited experiments we tried.