Hierarchical Part-based Disentanglement of Pose and Appearance

Project description: Learning image representations is a crucial task in Computer Vision. But common methods for this task lack interpretability, providing representation for an image as a whole, in a way that we don’t know what information has been encoded in each part of the represenation vector. Our goal in this project is to learn interpretable represantations for images without any supervision. We are specifically interested in encoding the parts that form an image, then for each part we learn its disentangled shape and appearance, and we also learn the hierarchy of detected parts. Learned representations by our model, are general purpose and could be used in many tasks such as Pose and Appearance Transfer, or Landmarks Detection.

For more details see the thesis, master’s presentation and source code.