Question about the number of disparity bins

[EmanueleGhelfi]

Hi,

Reading your paper, I found this sentence:

D_k \in  { Dmax/4, Dmax/2, Dmax, Dmax} represent the number of feature channel and the number of disparity bins
in the kth scale.

From this sentence, it seems that the number of disparity bins is exactly the max_disparity parameter passed to the network.

After inspecting the code in your repository, I found that the number of disparity bins is max_disparity / 16 for the finer resolution output.

Is the implementation different from the paper description? Did I interpret wrongly the description?

Have you tried to use a number of bins > max_disparity / 16 to increase the disparity resolution?

Thank you!