A novel 2D/3D database with automatic face annotation for head tracking and pose estimation

摘要

A new public database of videos for head tracking and pose estimation is presented in this paper with the goal of establishing a new framework for algorithm validation, replacing out of date frameworks. Position data has been recorded with a magnetic sensor-transmitter that has previously been aligned and synchronized with a commercial webcam, and we provide reliable ground-truth for 3D rotation and translation of the head with respect to the camera. In addition to this, an automatic face annotation procedure has been developed, which provides the image position of 54 facial landmarks, with negligible error, in every video frame in the database. This image ground-truth can be used for algorithm training or head tracking evaluation, among others. In order to show the usability of the database, we evaluate three head tracking approaches and three head models, and combine them to provide nine different head pose estimation sets of results. We show the validity of the presented database both for training and evaluation of head tracking and pose estimation methods, and provide an interesting comparison in performance of state-of-the-art algorithms. These results may also serve as reference to encourage other researchers to train and test their algorithms with this database, and compare their results with the ones presented in this paper.