Gates for handling occlusion in Bayesian models of images

Probabilistic systems for image analysis have enjoyed increasing popularity within the last few decades, yet principled approaches to incorporating occlusion as a feature into such systems are still few [11,10,7]. We present an approach which is strongly influenced by the work on noisy-or generative factor models (see e.g. [3]). We show how the intractability of the hidden variable posterior of noisy-or models can be (conditionally) lifted by introducing gates on the input combined with a sparsifying prior, allowing for the application of standard inference procedures. We demonstrate the feasibility of our approach on a computer vision toy problem.