Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 224

integration of sparse depth information, we have suggested an optimisation scheme
that simultaneously adapts the surface gradients to the measured intensity and po-
larisation data and to the surface slopes implied by depth differences between pairs
of depth points. This approach transforms sparse depth data into dense depth differ-
ence data, leading to a non-local influence of the depth points on the reconstructed
surface profile.
The experimental evaluation based on synthetic data illustrates that including
polarisation information into the three-dimensional reconstruction scheme signifi-
cantly increases the accuracy of the reconstructed surface. The main benefit arises
from taking into account polarisation angle data, while intensity and polarisation
degree tend to contain redundant information. We found that taking into account
dense but noisy depth from defocus data may be helpful in estimating the surface
albedo and in avoiding local minima of the error function. The integration of sparse
but accurate depth data significantly increases the three-dimensional reconstruction
accuracy, especially on large scales.
5.4 Stereo Image Analysis of Non-Lambertian Surfaces
This section describes a framework for the stereo analysis of non-Lambertian sur-
faces. The presentation is adopted from Wöhler and d'Angelo ( 2009 ). A general
drawback of all methods for stereo image analysis mentioned in Sect. 1.5 is the
fact that they implicitly assume Lambertian reflectance properties of the object sur-
faces. Two images of the surface are acquired from different viewpoints, and two
image parts are assumed to correspond to the same physical surface point if their
appearances are similar. This is only the case for Lambertian surfaces, where the
surface brightness is independent of the viewpoint. However, even in the Lam-
bertian case geometric distortions between the two images occur, which may be
taken into account by estimating an affine transformation between the views (Shi
and Tomasi, 1994 ). A method for three-dimensional reconstruction of objects from
image sequences that accounts for the changing camera viewpoint by a combination
of structure from motion and photometric stereo is presented by Lim et al. ( 2005 ).
The reflectance behaviour of the surface, however, is still explicitly assumed to be
Lambertian.
Since specular reflections are viewpoint dependent, they may cause large inten-
sity differences at corresponding image points. As a consequence of this behaviour,
stereo analysis is often unable to establish correspondences at all, or the inferred
disparity values tend to be inaccurate, or the established correspondences do not
belong to the same physical surface point. Only a few methods that specifically
address the three-dimensional reconstruction of specular surfaces have been pro-
posed in the literature. Lowitzsch et al. ( 2005 ) introduce a technique based on de-
flectometry, i.e. projection of light patterns on the surface and analysis of the de-
formations of their reflections, which is restricted to mirror-like surfaces. Bhat and
Nayar ( 1998 ) infer a relationship between the relative angle between the optical