Check for the idea of Segcaps

[ierosodin]

First, I think it's a great to extend the idea of convolutional capsules with locally-connected routing and propose the concept of deconvolutional capsules. I want to check the implementation of your work.
After applying transformation (maybe with a kernel size of 3 x 3), we route across different input capsules with a kernel size of "1 x 1", right? Or if we look transformation part together, we assume that the routing distribution of the input capsules are the same across different location within the kernel size so that we summation them (transformation part) and just do a 1 x 1 kernel size routing.
Please correct me if there is any wrong, thanks!