nal is oriented
as a result the
ad to optimize
uilding are not
len faces. The
ypical imagery.
image and shadow
vasis for many 3-D
form hexagons and
ry surfaces. Finally,
e used to generate
and Sander, 1992,
information sources
irface and its mate-
sses of information.
ion sources that do
s texture gradients,
ing multiple images
owever, we can do
single image. First,
r consistency across
. Second, when the
taken into account,
he images, thereby
tion.
nation sources that
iangulation of corre-
iven camera models
peaking, this source
can be easily identi-
measured. The ease
n vary significantly
depend critically on
vhatever the type of
fy where in the im-
ondences, and what
for correspondence
a possible) is simply
res—for example, by
reliable feature only
on the surface and,
textured.
often use fixed-size
(f)
Figure 5: Buildings modeled by entering rough models within RCDE and optimizing them using the extruded snakes. (a)
Rough initial sketches overlaid on one of the images. (b) A view from a different perspective. (c,d,e) Final building
outlines overlaid on the three images we used to perform the 3-D optimization. (f) A view of the buildings from
the perspective of (b).
windows in images to measure disparities, which will in gen-
eral yield correct results only when the surface is parallel to
the image plane. Instead, we compare the intensities as pro-
jected onto the facets of the surface. Consequently, the re-
construction can be significantly more accurate for slanted
surfaces. Some correlation-based algorithms achieve simi-
lar results by using variable-shaped windows in the images
[Quam, 1984, Nishihara, 1984, Kanade and Okutomi, 1990,
Baltsavias, 1991, Devernay and Faugeras, 1994]. However,
they typically use only image-centered representations of the
surface.
Our approach is much more closely related to the least-
squares approaches advocated by Wrobel [1991] and
Heipke [1992], who both use a 2-1/2-D representation of
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
the surface.
As for the monocular information source, we have chosen to
use shading, where shading is the change in image intensity
due to the orientation of the surface relative to a light source.
We use this method because shading is most reliable when
the albedo varies slowly across the surface; this is the natu-
ral complement to intensity correspondence, which requires
quickly varying albedo. The complementary nature of these
two sources allows us to accurately recover the surface ge-
ometry and material properties for a wide variety of images.
In contrast to our approach, traditional uses of shading in-
formation assume that the albedo is constant across the en-
tire surface, which is a major limitation when applied to real
images. We overcome this limitation by improving upon a
MEME
———— ET
i
