Avoiding the “Streetlight Effect” Tracking by Exploring Likelihood Modes.pdf

Avoiding the “Streetlight Effect”: Tracking by Exploring Likelihood Modes ICCV 2005 Submission 1614 Abstract Ideally we would like to evaluate the likelihood of a very broad and dense set of samples from the prior but in prac- Classic methods for Bayesian inference effectively con- tice this is impractical with existing probabilistic filtering strain search to lie within regions of significant probability methods. Broad search requires an extremely large num- of the instantaneous temporal prior. This is efficient with ber of samples, which are too costly to test and propagate an accurate dynamics model, but otherwise is prone to ig- individually. However, with a sharp likelihood and a wide nore significant peaks in the true posterior. A more accu- prior the shape of the posterior distribution depends much rate posterior can be obtained by explicitly finding modes more on the shape of the likelihood than on the temporal of the likelihood function and combining them with an ap- prior. Tracking performance may thus be improved by find- proximate temporal prior. In our approach modes are found ing modes of the likelihood function first and incorporating using efficient example-based matching followed by local prior information later. refinement to find peaks and estimate peak bandwidth. By In this paper we show how a broad search for modes of reweighting these peaks according to the temporal prior we the likelihood function can proceed efficiently, mitigating obtain an estimate of the full posterior model. We show the streetlight effect by considering regions of state space comparative results on real and synthetic images in a hi