Automatic Personal Annotation from Head Pose in Multi-View Systems

We developed an algorithm for automatic personal annotation for multi-view systems. Several problems arise when the annotation process is implemented in such a system. One of these problems is connected to the credibility criterion – the selection of optimal head poses. Another problem is connected to the number of images sufficient for the multi-view system to recognize the individual. An additional challenge is how to organize and store the extracted face images. Our algorithm applies a combination of CNN and SVM for the annotation. With the help of a Kinect sensor, the person is detected and his face is localized. The algorithm annotates the poses on the basis of the three angles of the head – yaw, pitch and roll. They are recorded in the database with the corresponding angles. The neural network was tested first on the CroppedYale B dataset and then on a dataset created by Kinect. After tuning the network’s parameters, almost all images with varying poses were classified correctly. Experiments were also made to find the range of angles, suited for the accurate annotation. With additional tests, the minimal number of images for individuals sufficient to train the network, was explored as well.