Adaptive Learning-based Temporal Tracker for 3D Head Shape Models

We propose a learning-based 3D temporal tracker that estimates the orientation and location of the head in the 3D scene. The algorithm is based on random forest that learns the 6D pose from a class of head models. A unique attribute of our approach is the capacity to adapt the learned tracker for a specific user, i.e., after learning, the tracker can deform the shape of the learned model to a specific instance of the class in order to match the user’s head shape. To find the user’s head shape model, we use a fast calibration method to personalize the model for a specific user. As a consequence, this technique enhances the accuracy of the head pose estimation as the personalized model becomes more detailed, and tracks at 1.4 ms per frame using a single CPU core.     «

We propose a learning-based 3D temporal tracker that estimates the orientation and location of the head in the 3D scene. The algorithm is based on random forest that learns the 6D pose from a class of head models. A unique attribute of our approach is the capacity to adapt the learned tracker for a specific user, i.e., after learning, the tracker can deform the shape of the learned model to a specific instance of the class in order to match the user’s head shape. To find the user’s head shape mo...     »