I have always had doubts about the necessity and intuitive/theoretical justification for capsule network in image classification and more recently nlp tasks. For the former, in order to address the global pose recognition problem, can’t we borrow idea from the LSTM/ attention to generalize the highly localized CNN by allowing more direct long Distance interactions between pixels far apart? Attention plus 2d position encoding seems a lot more natural and simpler than the elaborate dynamic routing approach advertised in the original capsule paper. Since I don’t work in the field of cv, this question may be naive, but really piques my mind and several others. Thanks!