Computer Perceptual Organization in Computer Vision

The following sections are included:

• Preface

• Contents

• List of Tables

• List of Figures

The following sections are included:

• What?

• Why?

The following sections are included:

• Classificatory Structure: A Proposal
  • 2D Organization
  • 3D Organization
  • 2D + Motion
  • 3D + Motion

• Computational Techniques
  • Group Theory
  • Coding Theory
  • Rule Based
  • Frequency Methods
  • Energy/Probabilistic Optimization Methods
  • Graph Theory

Perceptual organization provides significant computational leverage and can do so over several layers of abstraction. The sophistication of a vision system, on the whole, lies largely in the sophistication of its perceptual organization processes. In biological systems, visual capability is developed to a degree appropriate to the ambulatory capabilities of the organism, i.e., the degree to which it can use the information to acquire food, elude danger, or otherwise restructure its environment. The performance of a machine vision system is coupled to the chosen problem domain, which is analogous. It is the recognition, development, and exploitation of perceptual organization concepts, models, paradigms, and computational techniques that brings efficiency to machine vision systems. Perceptual organization allows us to assign computational resources intelligently, which is also important in biological systems because of the relative expense of neural tissue from an evolutionary standpoint. As we shall see, perceptual organization in computer vision uses computational resources effectively to extract organizations from which features are hypothesized, instead of the expensive alternative of directly applying feature detectors over the entire image.

The following sections are included:

• The Gestalt Graphs
  • First Level Organization
  • Second Level Algorithm

• Voting Scheme
  • Offset Quantization
  • Handling General Compatibility Relations
  • Complexity Analysis
  • This is Not a Hough Transform

• Selecting the Organization Parameters

• Result on the Aerial Image

The following sections are included:

• Knowledge Base: The Perceptual Inference Network
  • Belief Revision Using Modified Message Passing
  • Network Instantiation
  • Prediction of structures
  • Management of Expensive Computational Resources

• An Example PIN for Perceptual Organization
  • Manually Structuring the PIN
  • Location Compatibility Functions, f_LOC
  • Conditional Belief Functions
  • Prior Probabilities
  • Evidence Instantiation

• Algorithmic Complexity

• Results
  • Aerial Image
  • Performance Evaluation

The following sections are included:

• Overview of the Algorithm

• A Structural Representation of Organization
  • The Random Parametric Structural Description
  • Dissimilarity Measure
  • Structure of the Merged RPSD,
  • Constructing the Mapping between RPSDs

• Structuring the PIN
  • Dependency Detection
  • Next Level Formation
  • Root Node Decomposition
  • Expected Network Size
  • Desired Conditions

• Estimating the Conditional Probabilities
  • Estimating P(X = 1|U₁ = u₁, ⋯, U_n = u_n)
  • Computing f_LOC(l_k, l_k1,⋯, l_kn)

• An Example
  • The Target and Elementary Organizations
  • The Primitive and Relation Sets
  • Execution Statistics
  • Robustness

• Performance on Real Images
  • Comparing Automatically and Manually Structured PINs
  • Performance on Real Data
  • Performance Evaluation

The following sections are included:

• Theory
  • Control Algorithm
  • Information Theoretic Formalism
  • Return Condition
  • Final Hypothesis Resolution

• Special Purpose Modules
  • Corner Analysis Module
  • Segment Analysis Module
  • Region Analysis Module

• Results

• Performance Evaluation
  • Methodology
  • Results

The following sections are included:

• One Level of Organization
  • Blocks World
  • The laboratory image
  • Room Image

• Two Levels of Organization Hierarchy
  • Runway Image
  • Aerial Stereopair

• Timing Considerations

We presented a theoretical framework to compute perceptual organization in the context of computer vision. It is built around voting methods, graph enumeration, probabilistic Bayesian Networks, and information theoretic resource management. It incorporates a number of concepts from human visual analysis especially the Gestalt laws of organization and the notion of preattentive and attentive vision. We test our algorithm on a set of images from various domains. The methods is also computationally efficient and is potentially highly parallelizable.

The following sections are included:

• Appendix A
  • Tutorial for Bayesian Networks
    • Bayesian network dependencies
    • Belief updating in networks

• References

• Index