A particularly simple, but surprisingly effective, way to incorporate facial geometry into the tracking procedure is by projecting the output of the feature detections onto the linear shape model's subspace. This amounts to minimizing the distance between the original points and its closest plausible shape that lies on the subspace. Thus, when the spatial noise in the feature detections is close to being Gaussian distributed, the projection yields the maximum likely solution. In practice, the distribution of detection errors on occasion does not follow a Gaussian distribution and additional mechanisms need to be introduced to account for this.