An Empirical Study of Gaussian Belief Propagation and Application in the Detection of F-formations
Francois Kamper · 2017
In this paper we conduct an empirical study of the application of node-regularization within the context of Gaussian belief propagation (GaBP),and discuss a novel clustering method for the detection of F-formations using the message-components of this algorithm. In this empirical study we show that node-regularization can substantially improve the convergence speed of GaBP (up to 34.76 times less iterations to converge on average) and provide more accurate approximations for the marginal precisions (up to 13.92 times more accurate based on KL-distance). We show that our F- formation algorithm can provide state of the art accuracy scores and holds certain advantages in terms of dynamic modeling.