Consistency and Coupling in Human Model Likelihoods

This paper presents two novel likelihood terms for silhouettes and contours in model-based contexts. Despite the power of such formulations, building likelihoods that truly reflect the good configurations of the problem is by no means easy due to, most commonly, the violation of consistency principle resulting in the introduction of spurious, unrelated peaks/minima that make target localization difficult. We introduce an entirely continuous formulation which enforces consistency by means of an attraction/explanation pair for silhouettes. For contours, we address the search window vs. noise level dilemma by means of a combined robust estimation and feature coupling solution which builds a likelihood model not only in terms of individual contour responses but also “Gestalt” type higher-order couplings between matched configurations. We subsequently show how the proposed method provides significant consolidation and improved attraction zone around the true cost minima and elimination of some of the spurious ones.