Paper Review #13: SketchREAD

1. Paper Bibliography

• Title: SketchREAD: A Multi-Domain Sketch Recognition Engine

• Authors: Alvarado, Christine, and Randall Davis.

• Publication: Proceedings of the 17th annual ACM symposium on User interface software and technology. ACM, 2004.

2. Summary

• Provide structure description of the shapes in a certain domain (No training data or programming is necessary.)
• Context
• Guide the search for possible interpretations
• Reclassify low-level shapes
• Effect:
• Reduce recognition errors 
• Robustness to ambiguity and uncertainty inherent in complex, freely-drawn sketches
• Bayesian network
• Use a novel form of dynamically constructed Bayesian networks to evaluate these interpretation
• Effect:
• Recover from low-level recognition level
• Target domains: family trees and circuit diagram

3. Terminology

• Interpretation
• Hypothesis

4. Details

• The description of a shape

• Recognition
• Problem
• Real-time system -> Noise -> low-level misinterpretation -> higher-level interpretation fail
• Guarantee all interpretation -> exponential cases
• Solution
• Bottom-up recognition
• Generate most likely interpretation first
• Top-down
• Seek out missing interpretations
• Bayesian Network
• Evaluate partial interpretation hypotheses 
• Expand the hypothesis space by exploring the most likely interpretation first
• Hypothesis Evaluation
• A directed acyclic graph
• Which factors influence each other
• A set of conditional probability distributions
• How these factors influence one another
• Hypothesis Generation
• Top-down, Bottom-up, Pruning
• Selecting an Interpretation

5. Evaluation

• Circuit diagram is longer than family tree.