We use the the normalized intensity values as representations for SVM with linear and polynomial kernels. We use the SvmFu package (available at http://five-percent-nation.mit.edu/SvmFu/) and MIT CBCL Face Data to train our classifiers. In our experiments, the parameter C has been set to 10.
The training set consists of 6,977 images (2,429 faces and 4,548 nonfaces), and the test set consists of 24,045 images (472 faces and 23,573 nonfaces). (We would like to thank MIT CBCL for distributing the face data set). Our training and test sets are similar to the ones used in some recent works which show that SVMs with the feature representation of normalized intensity values perform better than the ones with Harr wavelet and gradient representations. In our experiment, each image is normalized to 20 by 20 pixels and processed with histogram equalization and quantization (50 rather than 256 scales). The following figure shows some face images in the training and test sets.
We train SVMs using linear and 2nd degree polynomial kernels. The empirical results on a test set of 24,045 images (472 faces and 23,573 nonfaces) are shown using the ROC (Receiver Operator Characteristic) curves.
