Face Detection
Current face recognition focuses on detecting tiny and heavily obscured faces. Facial analysis methods, however, require good localization and would benefit greatly from rotation information. We propose to predict a face direction vector (FDV) that provides face size and orientation and can be better learned from a common object recognition architecture than the traditional bounding box. It offers a more consistent definition of the position and size of the face. The use of the FDV shows promise for all subsequent facial analysis methods. As an example, we show that facial landmark recognition can benefit greatly from pre-aligned faces.
Example for detection of arbitrarily rotated faces. Blue: classic bounding box, red: suggested detection method.
Publications
Saxen, F., Handrich, S., Werner, P., Othman, E., Al-Hamadi, A., 2019. Detecting Arbitrarily Rotated Faces for Face Analysis, in: IEEE International Conference on Image Processing (ICIP). IEEE, Taipei, Taiwan.