Integrated Frameworks for Three-Dimensional Scene Reconstruction - 3D Computer Vision: Efficient Methods and Applications - page 225

axes of a pair of stereo cameras, the roughness of the surface, and the probabil-
ity of the correctness of an established stereo correspondence. This approach at-
tempts to minimise the influence of specular reflections on the images by choos-
ing an appropriate configuration of the stereo camera pair but does not quantita-
tively consider the effect of specular reflections for establishing stereo correspon-
dences.
A method for separating the specular from the Lambertian reflectance component
based on the 'dichromatic reflection model' is described by Klette et al. ( 1999 ). The
surface material is assumed to be dielectric and to consist of a coloured medium and
an interface. Body reflection, i.e. reflection of incident light at the optically neutral
medium, is diffuse, while reflection at the interface is specular. The resulting surface
radiance is characterised by two components, the 'interface reflection colour' and
the 'body reflection colour'. Klette et al. ( 1999 ) show that the RGB colour values of
the light reflected from the surface lie in the 'dichromatic plane'. Clustering in this
plane allows one to construct an image which only displays the diffuse reflectance
component, which can then be used for three-dimensional surface reconstruction
based on shape from shading or photometric stereo.
Lohse et al. ( 2006 ) propose the 'multi-image shape from shading' method, which
can in principle be used with an arbitrary but precisely known reflectance function.
A three-dimensional reconstruction of the surface is performed directly in the world
coordinate system along with an estimation of the reflectance parameters based on
an optimisation procedure, where a uniform surface albedo is assumed. This method
is favourably used under an oblique viewing geometry since under such conditions
small depth differences more strongly translate into offsets in the image coordinates
and thus have a more pronounced influence on the utilised error function. How-
ever, the experimental evaluation by Lohse et al. ( 2006 ) does not involve specular
surfaces.
Stereo image analysis which is independent of the reflectance function can be
achieved by taking into account the Helmholtz reciprocity condition (cf. ( 3.7 )),
which states that the value of the bidirectional reflectance distribution function
(BRDF) does not change upon exchange of the camera and the light source
(Magda et al., 2001 ; Zickler et al., 2002 , 2003a , 2003b ). This approach, how-
ever, requires a considerable instrumental effort. The Helmholtz stereopsis method
introduced by Zickler et al. ( 2002 ) consists of combining each camera with a
point light source in order to acquire 'reciprocal image pairs', such that the im-
age taken by the first camera is illuminated by the second light source and vice
versa. As a consequence, identical radiances are received by the first and the sec-
ond camera, respectively, where Zickler et al. ( 2002 ) point out that one pair of
images is generally not sufficient to obtain a three-dimensional scene reconstruc-
tion.
Other methods for BRDF independent stereo image analysis are based on vari-
ations of the illumination intensity; i.e. the light sources are kept at fixed locations
while the emitted intensity distribution is changed. This is the case e.g. for the pro-
jection of light patterns, termed 'coded structured light' by Batlle et al. ( 1998 ),
and also the spacetime stereo framework introduced by Davis et al. ( 2005 )(cf.