This repo is an implementation of Geoff Hinton's CapsNet in Pytorch. In addition, it contains an implementation of a ConvNet similar in structure to the CapsNet but without squashing or dynamic routing. These networks are compared by evaluating them on a dataset generated by transforming logos with random homographies. I found that the CapsNet was able to learn a model that was much more robust to homography transformations than the ConvNet was.
This project was created as my Computer Vision semester project, for which the writeup is in report.pdf.