A stochastic graph grammar for compositional object representation and recognition

  • Authors:

  • Liang Lin;Tianfu Wu;Jake Porway;Zijian Xu

  • Affiliations:

  • Beijing Institute of Technology, Beijing 100081, PR China and Lotus Hill Research Institute, Ezhou 436000, PR China and Department of Statistics, University of California, Los Angeles, USA;Lotus Hill Research Institute, Ezhou 436000, PR China and Department of Statistics, University of California, Los Angeles, USA;Department of Statistics, University of California, Los Angeles, USA;Department of Statistics, University of California, Los Angeles, USA

  • Venue:
  • Pattern Recognition
  • Year:
  • 2009

Quantified Score

Hi-index 0.01

Visualization

Abstract

This paper illustrates a hierarchical generative model for representing and recognizing compositional object categories with large intra-category variance. In this model, objects are broken into their constituent parts and the variability of configurations and relationships between these parts are modeled by stochastic attribute graph grammars, which are embedded in an And-Or graph for each compositional object category. It combines the power of a stochastic context free grammar (SCFG) to express the variability of part configurations, and a Markov random field (MRF) to represent the pictorial spatial relationships between these parts. As a generative model, different object instances of a category can be realized as a traversal through the And-Or graph to arrive at a valid configuration (like a valid sentence in language, by analogy). The inference/recognition procedure is intimately tied to the structure of the model and follows a probabilistic formulation consisting of bottom-up detection steps for the parts, which in turn recursively activate the grammar rules for top-down verification and searches for missing parts. We present experiments comparing our results to state of art methods and demonstrate the potential of our proposed framework on compositional objects with cluttered backgrounds using training and testing data from the public Lotus Hill and Caltech datasets.