Interpretation of Complex Scenes Using Generative Dynamic-Structure Models

Abstract

We propose a generative modeling framework - namely, Dynamic Tree Structured Belief Networks (DTSBNs) and a novel Structured Variational Approximation (SVA) inference algorithm for DTSBNs - as a viable solution to object recognition in images with partially occluded object appearances. We show that it is possible to assign physical meaning to DTSBN structures, such that root nodes model whole objects, while parent-child connections encode component-subcomponent relationships. Therefore, within the DTSBN framework, the treatment and recognition of object parts requires no additional training, but merely a particular interpretation of the tree/subtree structure. As such, DTSBNs naturally allow for multi-stage object recognition, in which initial recognition of object parts induces recognition of objects as a whole. As our reported experiments show, this explicit, multi-stage treatment of occlusion outperforms more traditional object-recognition approaches, which typically fail to account for occlusion in any principled or unified manner.