Applications to Industrial Quality Inspection - 3D Computer Vision: Efficient Methods and Applications - page 275

the surface measured with the SfPRD method is about twice as high as the lateral
pixel resolution of the utilised images. To obtain depth from defocus information,
the acquisition of two images at different apertures may be automated using a motor
iris lens.
Large-scale deviations of the estimated three-dimensional surface shape from the
true shape may occur if the available depth data are too sparse. Hence, if traditional
stereo analysis techniques do not yield three-dimensional point data of satisfactory
accuracy or density, the specular stereo method suggested by Wöhler and d'Angelo
( 2009 ) (cf. Sect. 6.3.4 ) establishes a number of stereo correspondences that may
become several orders of magnitude higher than the number of correspondences ob-
tained by classical blockmatching. The three-dimensional point cloud obtained with
the specular stereo method is less noisy, contains a negligible number of outliers,
and shows significantly more surface detail than the point cloud obtained by classi-
cal blockmatching. For poorly known reflectance parameters, a graceful degradation
of the accuracy of the three-dimensional reconstruction result is observed.
The approach proposed by Herbort et al. ( 2011 ) (cf. Sect. 6.3.5 ), which combines
active range scanning data with photometric stereo, yields a three-dimensional re-
construction result which is accurate on both large and small spatial scales. Most
details visible in the acquired images are also apparent in the reconstructed depth
map, while the strong noise in the active range scanning data is removed and the
gaps are filled. The accuracy of the three-dimensional surface reconstruction result
is about twice as high as the lateral pixel resolution of the images.