Skip to Main Content
Head pose estimation, though a trivial task for the human visual system, remains a challenging problem for computer vision systems. The task requires identifying the modes of image variance that directly pertain to pose changes, while generalizing across face identity and mitigating other image variances. Conventional methods such as Principal Component Analysis (PCA) fail to identify the true relationship between the observed space and the pose variable, while supervised methods such as Linear Discriminant Analysis (LDA) neglect the continuous nature of pose variation and take a discrete multi-class approach. We present a method for estimating head pose using Canonical Correlation Analysis (CCA), where pose variation is regarded as a continuous variable and is represented by a manifold in feature space. The proposed technique directly identifies the underlying dimension that maximizes correlation between the observed image and pose variable. It is shown to increase estimation accuracy and provide a more compact image representation that better captures pose features. Additionally, an enhanced version of the system is proposed that utilizes Gabor filters for providing pose sensitive input to the correlation based system. The preprocessed input serves to increase the overall accuracy of the pose estimation system. The accuracy of the techniques is evaluated using the Pointing '04 and CUbiC FacePix(30) pose varying face databases and is shown to produce a lower estimation error when compared to both PCA and LDA based methods.