Cascading Deformable Parts Model in the Facial

Feature Detection of Frontal and Side Profile Images

Pisal Setthawong and Vajirasak Vanijja

School of Information Technology, King Mongkut's University of Technology, Thailand

Abstract: Facial feature detection is considered an important computer vision task that is used for many real world applications. Current advancements in computer vision have come up with many proposed facial feature detection approaches, such as Deformable Parts Model (DPM), that provide good accuracy in the detection of key facial features, but mainly in frontal poses. When presented with side profile poses many approaches do not perform well as certain facial features can be obscured and the approach may attempt to over-fit the trained model which leads to inadequate results. The proposal of a cascading pipeline extension to the DPM approach and a modified DPM approach for side profile specific facial feature detection is presented to deal with a wider range of facial profiles. The proposed approach would be evaluated empirically showing the improvement of the proposed method over a wide range facial feature configurations including side and frontal profiles.

Keywords: Facial feature detection, DPM, geometric model, image processing.