A Methodology for Designing Computer Vision Systems


The PCCV Project:
Benchmarking Vision Systems

Case studies
Test datasets
Our image file format
HATE test harness

General links
Mailing lists
Research groups

Techniques (CVonline)
Image databases

Other stuff
Linux on ThinkPad

Some work has been carried out under the PCCV programme into a statistically-valid methodology for designing and constructing computer vision systems. This is summarized in the following document:

An Empirical Design Methology for the Construction of Computer Vision Systems: 32pp report, available as PDF (268 Kbytes), PostScript (2.3 Mbytes)

ABSTRACT: The main body of this document covers the essential foundations of design methodology for machine vision algorithms, making explicit the links with conven tional statistical principles. We explain the necessary link between design and testing. Numerous practical problems in the analysis of noisy data are addressed while an extensive set of appendices give technical detail which briefly describe techniques which have been used in computer vision algori thms to address problems such as error estimation, lack of data independence and data fusion. The techniques are illustrated with simple examples. For the particular issue of vision module construction and testing (which we call technology evaluation) we provide a flow chart showing how the techniques described in the appendices can be brought to bear at different s tages of the algorithm design process. We finish by suggesting the criteria by which empirical work in this area should be assessed if published results are to be used by others in the construction of larger systems.

Comments on this document from prominent computer vision researchers and our responses to them are also provided so that this work can be set in the context of current attitudes in the field of computer vision.

Last updated on 13-Jan-2004 by Adrian F. Clark. [Comment on this page]