Image Processing Reference
this was extended to reconstruction of deformable and articulated objects seen under orthographic
projection. Other methods were proposed in Del Bue et al. [ 2010 ], Wang et al. [ 2008 ]toallowfor
different camera models, as well as to improve the convergence properties of the original algorithm.
The above-mentioned techniques were designed to cope with missing data. A different prob-
lem arises from the presence of mismatches in the point tracks. To overcome these outliers, the most
common scheme is to rely on a RANSAC Fischler and Bolles [ 1981 ] procedure Olsen and Bartoli
[ 2008 ], Zhu et al. [ 2010 ]. Another solution involves using a robust estimator to weight the points
according to the uncertainty of the measurements Shaji and Chandran [ 2008 ]. This technique also
deals with missing data by assigning a zero weight to the lost points.
In any event, even though there are techniques designed to overcome the missing data prob-
lem, the theoretical solutions discussed in this chapter are not sufficiently constrained to recover
meaningful structure and motion by themselves. As for template-based reconstruction, additional
knowledge needs to be introduced in NRSFM algorithms. In the next chapter, we will discuss the
different types of knowledge that have been proposed so far.