Safari Books Online is a digital library providing on-demand subscription access to thousands of learning resources.
- Create BookmarkCreate Bookmark
- Create Note or TagCreate Note or Tag
- DownloadDownload
- PrintPrint
Chapter 6. Feature Selection in High Dimension > EXPERIMENTS AND CONCLUSION
EXPERIMENTS AND CONCLUSION
Dataset Description and Objectives
We have designed two sequences of experiments for the problem of selecting variables for the detection of "face" versus "non-face".
We use first the AT&T ORL (Olivetti) database composed of 400 samples of 92 × 112 faces with 256 grey levels per pixel. Our features are the Sobel vertical and horizontal edge detectors, this yields a dimension of 20608. As no test set is available, we split the data in two parts. We use for non faces samples of images of the same size taken on the Web. Some samples extracted from the database are shown in Figure 1.
We use also the MIT-CBCL database that contains face images of 19 × 19 pixels. These images are described with the binary edge detectors of Fleuret and Geman (2001). We then obtain 1926 binary features defined by their orientation, vagueness and position. The database provides both learning and test set, each one described with face and non face images. Some samples extracted from the database are shown in Figure 2.
