PROJECT TITLE :

Head Pose Estimation Based on Multivariate Label Distribution

ABSTRACT:

The training of the majority of currently available head pose estimation methods requires accurate ground-truth poses as a prerequisite. However, in many instances, the "ground truth" pose is obtained through rather subjective methods. For example, the subjects may be asked to stare at various markers on the wall in order to obtain this pose. To indicate the pose of a face image, it is therefore preferable to use implicit soft labels as opposed to explicit hard labels. This paper makes the suggestion that each image should be accompanied by a multivariate label distribution, or MLD. An MLD investigates the area immediately surrounding the original pose. Labeling the images with MLD can not only solve the issue of incorrect pose labels, but it can also increase the number of training examples that are associated with each pose without actually increasing the overall number of training examples. The MLD learning problem is addressed by four different algorithms. In addition, an extension of MLD called hierarchical multivariate label distribution has been proposed to deal with fine-grained head pose estimation. This extension is named after the hierarchical structure that it adds (HMLD). The experimental findings indicate that the MLD-based methods perform noticeably better than the state-of-the-art head pose estimation algorithms that were evaluated for comparison. In addition, in comparison to the baseline methods, the MLD-based methods appear to be significantly more resistant to the label noise that is present in the training set.