In supervised learning of head pose classification, uniformly distributed and labeled ground-truth of people in large quantities is required. Unfortunately, labeling is both tedious and inconsistent. As a result, the classifier could not generalize well the unseen data. To address this problem, in this paper we propose a novel bootstrapping (semi-supervised) which can train a classifier using both a small number of labeled data and an abundance of unlabeled data. The difficulty of using unlabeled data is that the performance could be worse than using just the labeled data if the unlabeled data has a different feature space distribution than that of the labeled data. This could be the reason why there is little work done to estimate the head pose with unlabeled data. In our proposed method, automatic data mining is applied to select unlabeled data that has higher likelihood to be helpful to improve the performance of a classifier trained solely on the labeled data. Kernel linear discriminant analysis and a SIFT-based image registration are combined to predict the head pose from face image. Some pose prototypes, learned from the labeled samples, are used to define a novel confidence measurement for selecting the unlabeled data. Experimental results on a large database verified that the proposed bootstrapped approach can achieve significantly better performance than the supervised learning alone.