Dynamic Positional Trees for Structural Image Analysis

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract / Description of output

Dynamic positional trees are a significant extension of dynamic trees, incorporating movable nodes. This addition makes sequence tracking viable within the model, but requires a new formulation to incorporate the prior over positions. The model is implemented using a structured variational procedure, and is illustrated on synthetic raytraced images and image sequences. We consider the problem of structural image analysis and in particular the inference of scene properties from image data. We are especially concerned with image decomposition, that is obtaining the characteristic parts of an image and the relationships between them. The components of an image are not independent of each other; certain objects are expected to occur together, and objects are made up of different subcomponents. One way of thinking of this problem is by analogy with parsing a language; we are interested in parsing images. However, the important characteristics and structure in an image is significantly different from linguistic data. Those familiar with work on dynamic trees will be aware that they have been developed in the context of single static images [15, 1, 13]. It would be desirable if the benefits of the dynamic tree approach could also be made available for image sequences. Introducing a sequence model into the basic dynamic tree formalism is not straightforward as a change in the position of an object is reflected in a change in the connectivity structure of the dynamic tree. This change would be hard to predict from the previous time slice and would be an inelegant representation of the dynamics: the connectivity structure is supposed to represent the structural characteristics of an object, most of which will be preserved during movement. Here the dynamic tree is modified to incorporate position variables, resulting in a model where object movement can be represented in terms of a change in position components of the nodes representing that object.
Original languageEnglish
Title of host publicationIn Proceedings of the Eighth International Workshop on Artificial Intelligence and Statistics
Pages298-304
Number of pages7
Publication statusPublished - 2001

Fingerprint

Dive into the research topics of 'Dynamic Positional Trees for Structural Image Analysis'. Together they form a unique fingerprint.

Cite this