Color Prior for Feature-based 3D Head Tracking

This paper proposes a novel color prior to reduce outlier keypoints for feature-based 3D head tracking. It is more effective than existing approaches for two reasons. First, the color information of a keypoint is represented by a powerful covariance descriptor (RGB Sigma Set). Second, the color prior is modeled from the posteriors of a discriminative classifier based on Random Projection trees. To demonstrate its strength, we integrate the prior into a 3D head tracking system that estimates pose by matching a set of wide-baseline keypoints. Our experiments show that most outliers are rejected by the proposed prior, which in turn significantly improves the accuracy, speed and robustness of the tracking system.