Abstract

We first address the question of how to build a ‘strong’ probabilistic model of object shapes (in the form of binary silhouettes). We define a ‘strong’ model as one which meets two requirements: 1. Realism – samples from the model look realistic, and 2. Generalization – the model generates samples that differ from training examples. We consider a class of models known as Deep Boltzmann Machines and show how a strong model of shape can be constructed using a specific form of DBM which we call the ‘Shape Boltzmann Machine’ (ShapeBM).

We also present a generative framework for modelling RGB images of objects. Building on the ShapeBM, our model employs a factored representation to reason about appearance and shape variability across datasets of images. Parts-based segmentations of objects are obtained simply by performing probabilistic inference in the proposed model. We apply the model to several challenging datasets which exhibit significant shape and appearance variability, and find that it obtains results that are comparable to the state-of-the-art.