The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9835

Title:Did the Earth Move for You ? A Novel Metric for Matching Iconic Images
Date: 1998
Abstract:The Iconic Vision System uses a combination of image-like and symbolic evidence to match images with models previously learned using a simple perception. Currently, matching employs a metric analogous to calculating the pixel-wise scalar product. Proposed is the use of a novel constrained version of "Earth Movers Metric (EMD)".The linear programming problem representing the EMD is optimised using the Simplex method and should give a measure with more flexibility. In the iconic system's R-theta representation, the allowed "shifts" between matchable elements represent local variations in scale and orientation.The constrained version of the EMD was developed to reduce the computational requirements and prevent unwanted "shifts" corresponding to large scale orientation variations.The curent tests indicate that while the approach has promise, use of the metric would slow the system down unacceptably. further work is rquired to speed up the computation, possibly by implementing a streamlined Simplex method and applying "early failure" conditions.Another new metric uses Gaussian blurring of the R-theta image to implement similar flexibility and seems to give comparable results with far less intensive computation.A secondary attempt to improve performance was retraining the perceptron used to weight contributions to the overall match from the range of 42 feature match scores. The results from the retraining were disappointing, and imply that the features currently used may be weak and possibly do not give linearly separable result sets.

[Search These Pages] [DAI Home Page] [Comment]