Detecting the dominant normal directions to the decision surface is an established technique for feature selection in high dimensional classification problems. Several approaches have been proposed to render this strategy more amenable to practice, but they still show a number of important shortcomings from a pragmatic point of view. This paper introduces a novel such approach, which combines the normal directions idea with Support Vector Machine classifiers. The two make a natural and powerful match, as SVs are located nearby, and fully describe the decision surfaces. The approach can be included elegantly into the training of performant classifiers from extensive datasets. The potential is corroborated by experiments, both on synthetic and real data, the latter on a face detection experiment. In this experiment we demonstrate how our approach can lead to a significant reduction of CPU-time, with neglectable loss of classification performance.