In this paper, we propose a novel unsupervised constellation model learning algorithm based on voting weight control for accurate scale, rotation, and translation invariant face localization without manual selection of feature points. The constellation model is learned by controlling the expected voting weights of the local features to obtain their perceptual boundaries and the distribution of voting weights, and selecting most common features as the representative features among them. The proposed constellation model can be learned incrementally to successfully localize faces when the previously learned model fails to localize them accurately. Through experiments, it is shown that the proposed constellation model can accurately localize faces of various size, orientation, and location. (C) 2008 Elsevier Ltd. All rights reserved.