Characterization of Algorithms for Passive Infrared (CAPIR).
Abstract
This study has evaluated three types of classifiers on a particular set of ATR data. The ATR data consisted of two sets of images: one used to train the classifiers and one used to test the classifiers. For each target object in the ATR data, nine simple shaped-based features were computed and this full nine-dimensional feature space (along with a few sub-spaces of smaller dimension) was used as the basis for classifier evaluation. The three classifiers chosen for evaluation were: a kernel-based classifier, the K-Nearest Neighbor (KNN) classifier, and a simple parametric Gaussian classifier. The error rate of each classifier was reported using the leave-one-out approach to error estimation over the training set and this error rate was compared to the actual error rate obtained on the test set. In addition, two further evaluations were performed. The first involved the selection of the window-width parameter needed in kernel-based density estimation. As well as the baseline leave-one-out method of setting this parameter, three additional selection procedures were evaluated: the MEISER technique, a Bhattacharya technique, and an density overlap technique. The second study involved the definition and evaluation of confidence measures for estimated densities. Two measures were evaluated: a bootstrap measure and a non-linear scaling measure.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 1987
- Accession Number
- ADA193069