Feature and Extractor Evaluation Concepts for Automatic Target Recognition (ATR)
Final rept. 1 Jan-1 Oct 1995
WRIGHT LAB WRIGHT-PATTERSON AFB OH
Pagination or Media Count:
This report develops concepts that will support the evaluation planning for the MSTAR features and feature extractors. These concepts will be used later in building a detailed evaluation plan. We began our development by distinguishing between the evaluation of a feature set and the evaluation of an extractor. The specifics for feature evaluation depend upon whether or not it is meaningful to define a truth-value but in either case, features are evaluated in terms of their sensitivity at first individually and then as a set to various factors. The factors of interest fall into the categories of Known, Class, and Noise. Ideal features would be discriminating high sensitivity to class factors, robust low sensitivity to noise factors, and predictable predictable sensitivity to known factors. The evaluation of extractors including auxiliary information such as runtimememory use estimates and feature uncertainty is based on accuracy when meaningful, design quality, and good software engineering principles.
- Target Direction, Range and Position Finding
- Active and Passive Radar Detection and Equipment