Junction trees (JT) is a general purpose tool for exact inference on graphical models. Many of the existing algorithms for building junction trees require a fixed static graphical model. The construction process is not unique, finding the one with the best computational structure (smallest clique size) is also a hard problem. For large scale inference problems, such as Geo‐referencing using triangular geodetic networks or equivalent, the simultaneous localization and mapping (SLAM) problem in robotics pose some challenges to junction tree applications.
Incremental junction tree techniques for dynamic graphical models prescribe heuristic methods for growing the tree structure, and are applicable to large scale graphical models. Of concern are the proliferative widening of the tree, which makes message passing expensive. In the context of SLAM we present a new apporach that exploits the local frame dependence of novel observation variables.
