There are a number of variables in the training and tracking process that can be tweaked to optimize the performance for a given application. However, one of the primary determinants of tracking quality is the range of shape and appearance variability the tracker has to model. As a case in point, consider the generic versus person-specific case. A generic model is trained using annotated data from multiple identities, expressions, lighting conditions, and other sources of variability. In contrast, person-specific models are trained specifically for a single individual. Thus, the amount of variability it needs to account for is far smaller. As a result, person-specific tracking is often more accurate than its generic counter part by a large magnitude.