In video situations, be it real-time or offline processing, as in the above cases corresponded features are needed for the SfM problem. However, in video, features are computed in a temporally incremental way or are 'tracked' through the sequence. Tracked features may thus exhibit significant noise levels which should be estimated and treated rigorously. A feature is simply a measurement and all measurements have error (even at the pixel level due to image digitization). Therefore, one should model the error, for example via a typical Gaussian distribution (which identifies an ellipsoidal iso-probability curve around 2D feature points). In addition, over a video sequence, this error is likely to vary over time when the feature gets occluded (i.e. generates a wide distribution) or is very clearly observable (i.e. generates a tight distribution). This information is clearly useful and hence should be included in an SfM framework.