Abstract

Images are composed as a hierarchy of object parts. We use this insight tocreate a generative graphical model that defines a hierarchical distributionover image parts. Typically, this leads to intractable inference due to loopsin the graph. We propose an alternative model structure, the Dense Latent Tree(DLT), which avoids loops and allows for efficient exact inference, whilemaintaining a dense connectivity between parts of the hierarchy. The usefulnessof DLTs is shown for the example task of image completion on partially observedMNIST and Fashion-MNIST data. We verify having successfully learned ahierarchical model of images by visualising its latent states.

 

Quick Read (beta)