<< Back

In this paper, we propose an approach for detecting facial features and recovering pose in presence of high out of plane rotations for both still images and video streams. To detect the correct features, we assign a confidence number to combinations of feature candidates given the edge map of the face. Feature candidates are determined using probability distribution of color space of skin, eyes and eyebrows. To increase the accuracy of feature detection for video streams, we incorporate motion history information for individual features by weighing the confidence measure according to potential regions of features. Once the best feature combination is obtained, we recover the pose using the centroid of the features assuming orthographic projection.

We conducted experiments on both still images and eleven video sequences including two CNN interviews. In almost all cases, the system performed very well and correctly determined the pose.

Associated publications:
Automatic Feature Detection and Pose Recovery for Faces
Asian Conference on Computer Vision, ACCV 2002, Melborne, Australia, Jan 2002