SIFT has also been applied in Photogrammetry, in close-range applications, for 3D modelling of small objects [20] and for spatio-temporal feature tracking analysis [21]. Moreover, SIFT has also been applied in remote sensing [22-23], in the registration of LIDAR intensity data and aerial images [24], in the co-registration of synthetic aperture radar interferometry data [25] and in real-time mapping applications from UAV [26]. Several methods similar to the SIFT operator method have been developed in order to overcome its high computational cost; however, faster implementations (PCA [27], SURF [28], etc.) reduce the point location accuracy.Although many papers and much research about feature detectors have been carried out within the CV community, detailed studies concerning the accuracy of the SIFT operator have never been performed in the Photogrammetry field.

Some articles which compare feature detectors can already be found in literature: Mikolajczyk [5-29] has analysed the performances of affine-invariant and scale invariant region detectors and Schmid [30] has evaluated the performances of interest point detectors. These papers evaluate the feature extractors in terms of the number of extracted points and repeatability and show that the SIFT detector supply more stable results than the other ones. However, the determination of the localization accuracy has only been performed on terrestrial images.Accuracy is the most important criterion for the evaluation of a good photogrammetric process.

For this reason, the main goal of researchers in photogrammetry is to assess the accuracy that feature points and region operators can reach in the automatic feature extraction and matching phases of the photogrammetric process. Remondino [31] has carried out tests on six regions and interest point detectors. He has compared the Batimastat results obtained from a quantitative analysis that was based on the relative orientation between image pairs. The test results, highlighted optimal performances of the region detectors (in particular SIFT) as far as the number of points extracted is concerned, even though the accuracy was not as high as that of the interest operator ones. Furthermore, the author showed that the accuracy of SIFT can be improved using the Least Square Matching (LSM) algorithm [32]. However, only a SIFT demo-version was dealt with in this paper and only terrestrial images were considered.

The performance analyses performed in the previous researches on the SIFT technique have dealt with the geometric and the illumination conditions of the image acquisition, but they did not consider the dynamic range of the image or the texture distribution. In [18] the importance of contrast thresholds of the SIFT in relation to the number of extracted points has been underlined.