Spatial Dependence in the Observation of Visual Contours

Two challenging problems in object recognition are: to output structures that can be interpreted statistically; and to degrade gracefully under occlusion. This paper proposes a new method for addressing both problems simultaneously. Specifically, a likelihood ratio termed the Markov discriminant is used to make statistical inferences about partially occluded objects. The Markov discriminant is based on a probabilistic model of occlusion which introduces spatial dependence between observations on the object boundary. This model is a Markov random field, which acts as the prior for Bayesian estimation of the posterior using Markov chain Monte Carlo (MCMC) simulation. The method takes as its starting point a “contour discriminant” designed to differentiate between a target and random background clutter.