Interpretation of Complex Scenes Using Dynamic Tree-Structure Belief Networks

In this paper, we address the problem of object detection and recognition in complex scenes, where objects are partially occluded. We speculate that a careful analysis of visible object details at various scales may prove critical for recognition in these settings. However, in general, computational complexity becomes prohibitive when trying to analyze multiple sub-parts of multiple objects in an image. To alleviate this problem, we propose a generative-model framework – namely, Dynamic Tree-Structure Belief Networks (DTSBNs) and a novel Structured Variational Approximation (SVA) inference algorithm for DTSBNs. Thus, we formulate object detection and recognition as inference of DTSBN structure and image-class conditional distributions, given an image. The causal (Markovian) dependencies in DTSBNs allow us to design computationally efficient inference, as well as to interpret the estimated structure as follows: each root represents a whole distinct object, while children nodes down the sub-tree represent parts of that object at various scales. 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. This property allows us to devise a strategy for recognition of objects as a whole through recognition of their visible parts. Our experimental results demonstrate that this approach remarkably outperforms strategies without explicit analysis of object parts.