Face Detection with Support Vector Machines and a Very Large Set of Linear Features

This paper presents a fast and novel method to speed up training and evaluation of support vector machine (SVM) classifiers with a very large set of linear features. A pre-computation step and a redefinition of the kernel function handle linear feature evaluation implicitly and thus result in a run-time complexity as if no linear features were evaluated at all. We then train a classifier for face detection on a set of 210,400 linear features. The resulting classifier has a support vector count and running time that is 50% lower than a classifier trained on raw pixel features, but still maintains a comparable classification performance. A recent feature weighting approach is adapted for large linear feature sets and used to improve the classification performance of both classifiers even further. Again, the classifier based on our linear feature set outperforms a comparable classifier based on pixel inputs.