Unsupervised Part Representation by Flow Capsules

     Sara Sabour 1 2 Andrea Tagliasacchi 1 2 Soroosh Yazdani 1 Geoffrey E. Hinton 1 2 David J. Fleet 1 2

             Abstract
    Capsule networks aim to parse images into a hi-
    erarchy of objects, parts and relations. While
    promising, they remain limited by an inability to
    learn effective low level part descriptions. To ad-
    dress this issue we propose a way to learn primary
    capsule encoders that detect atomic parts from a
...