Bayesian Inferences for Object Detection

Moghaddam et al. (Moghaddam and Pentland, 1997) first proposed to perform object detection by modelling the marginal probability density function (pdf) of high dimensional features of appearance. Based on gaussian hypotheses, their approach is however not robust to outliers that can occur in images due, for instance, to cluttered backgrounds or partial occlusions. In (Dahyot et al., 2004), robustness has been improved by using better priors for the distribution of the errors encountered in the observations. Because the marginal pdf was not analytically available, the Maximum A Posteriori (MAP) pdf was used instead for detection (Dahyot et al., 2004). If both pdfs, marginal and MAP, are proportional under gaussian assumptions (MacKay, 1995), and therefore perform equivalently, we aim in this paper to compare both densities (marginal and MAP) for object detection using the same robust priors.