In the last few years 3D face recognition has become more and more popular due to reducing cost of scanners and increasing computational power. The crucial and time-consuming step is landmark localization and normalization of facial surface. Due to acquisition, noise and other artifacts like spikes and holes occur. Most systems require computational intensive preprocessing steps to eliminate these artifacts. As a consequence, a trade-off between runtime or detection accuracy must be made. In contrast, we propose a landmark detection algorithm which uses the Viola \& Jones classifier on gradient images. The algorithm is able to reliably detect landmarks in raw 3D data without complicated preprocessing. Additionally, selection of sub regions is exploited to limit search regions. It further reduces false detection rate and improves significantly detection accuracy.