The Role of Middle Level Features for Robust Shape Tracking

Shape trackers using low level features (e.g., edge points) often fail in complex environments (e.g. clutter, inner edges or multi-objects). Two alternatives are discussed in this paper. Both methods use middle level features: (data centroids, strokes), which are more informative and reliable than edge transitions used in most shape trackers. Furthermore, it is assumed that each middle level feature is either valid or an outlier. Therefore a confidence degree is assigned to each feature. Features with a high degree of confidence have a large influence on the shape estimate while features with low degree of confidence have a negligeable influence on the final estimates. Both mechanisms (the use of middle level features and confidence degree) lead to a significant improvement of the tracker robustness. This is shown in the paper in the context of lip tracking problem.