In this work, we focus on the task of learning and representing dense
correspondences in deformable object categories. While this problem has been
considered before, solutions so far have been rather ad-hoc for specific object
types (i.e., humans), often with significant manual work in