Securing SIFT with A Contrario Techniques

In the previous chapter two shortcomings of Lowe's SIFT algorithm have been pointed out, namely its low matching efficiency (ratio between the number of correct matches and the total number of matches) and its inability to match several instances of the same object. The grouping stage of the method also is widely empirical and requires some fix.

In this chapter we shall examine three easy improvements of the SIFT method, all based on the a contrario techniques developed in the present book. They permit to treat all raised issues. The first one (Sect. 11.1) is the direct application of the theory for a contrario grouping of transformations developed in Chap. 8. The second one (Sect. 11.2) is the use of a background model for SIFT matches which prevents the elimination of multiple matches. Finally Sect. 11.4 yields an efficient a contrario technique computing a NFA for each SIFT match. In summary, the aim is to demonstrate that the whole SIFT algorithm can be secured and associated realistic NFAs, as we did in Chap. 5 and 8 for the LLD method.