The Role of Task Specification in Optimal Cue Integration
Paul Schrater, Daniel Kersten · 1999
Abstract. A full Bayesian approach to vision requires consideration of potential interactions between all the variables in both the scene and image. A complete model of the interactions, however, seems computa-tionally intractable because of the high dimensionality of image measurements and scene properties. As a consequence, both artificial visual systems and models of human vision have relied on a premature commitment to computationally simple modular architectures in which a particular scene property (e.g. object depth) is estimated from a restricted set of image measurements (e.g. image size), without reference to other scene properties. Such “Depth-from-X ” approaches create problems for inference that include invalid and contradictory assumptions and difficulties in fusing the potentially inconsistent estimates. The computational problem posed by optimal inference is not hopeless, however, and can be surmounted by restricting inference to particular tasks and taking advantage of the statistical structure of the problem. In a Bayesian context, modularity falls out of the conditional independencies in the joint distribution of scenes and images p(S, I). However, restricting optimal inference to particular tasks has the effect of making the choice of representation for the scene variables non-trivial. In particular, some representations lead to simpler inference algorithms than others. We illustrate the problem of modularity and cue combination