398 E. Guillou et al.: Using vanishing points for camera calibration and coarse 3D reconstruction from a single image
2.2 3D Reconstruction from a single image
or a single panoramic image mosaic
To our knowledge, four papers address the problem
of 3D reconstructionfrom a single image [14,16–18]
and one from a single panoramic image mosaic [15].
The method, described in [14], does not employ any
conventional computer vision technique, but makes
use of a spidery mesh to obtain a very simple scene
model from a single image using a user interface.
The aim is to provide a new type of visual effects
for various animations rather than to recover a 3D
model for a walkthrough. The method starts by spec-
ifying one vanishing point in the image. Then the
geometry of the background is approximated with
a model composed of five rectangles. This model is
a polyhedronlike form with the vanishing point on
its base. Each foreground object (near the viewer)
is modeled with hierarchical polygons. The bill-
boardlike representation is used for the foreground
objects.
Parodi and Piccioli [16] have investigated the recon-
struction of 3D models in a scene from line draw-
ing (on a single image). Geometrical constraints are
used,and they are providedby the location of vanish-
ing points and an incidence structure I represented
by a graph with nodes that are vertices and pla-
nar panels. A spatial structure S = (I, T) is also de-
termined, where T is a set of depth relations. The
connected components of this graph are determined.
A component can be seen as a different object or
a set of physically connected objects. Given the lo-
cation of a 3D vertex, the authors propose a formula
that determines the 3D coordinates of other vertices
belonging to the same planar panel. This formula
contains some terms depending on two vanishing
points determined from the planarpanel. The method
computes the vanishing points associated with each
planar panel. The focal length is calculated with the
algorithm described in [20]. A scale factor is associ-
ated with each component. To check the realizability
of the reconstruction, three independent tests are per-
formed in sequence. In addition, one of these tests
checks if there is a set of scale factors such that the
resulting 3D locations of all vertices simultaneously
satisfy the incidence and depth relations.
In [15], Shum et al. present an interactive model-
ing system that constructs a 3D model from one or
several panoramic image mosaics. The system re-
covers the camera pose for each mosaic from known
line directions and points, and then constructs the
3D model using all available geometrical constraints
(soft or hard). The modeling process is formulated as
a linearly constrained least-squares problem which
can be solved with the QR factorization method. The
method makes use of parallel lines and/or known
points to recover the camera rotation. For a single
panorama, the translation is set to zero if no point in
the 3D model is given. Textures are also extracted by
the method in [24].
Liebowitz et al. [17] propose a two-pass reconstruc-
tion method applied to planar façades. In a first step,
the visible façades are individually reconstructed by
rectification. The second step proceeds in an incre-
mental fashion, that is to say, the façades are posi-
tioned in turn. More precisely, a façade A can be
positioned relatively to another facade B only if A
and B share common points. The orientation of the
facade B, as well as its size, is determined with the
common points and the rectification parameters of
the planes supporting the two façades A and B.
Van Den Heuvel [18] describes a method for recon-
structing polyhedral object models from a single im-
age. The method uses measurements of image lines,
object information, etc. Topology (of the coplanarity
type) and geometry (parallelism, perpendicularity,
symmetry) constraints are used by the reconstruc-
tion process. The method assumes that the camera is
calibrated.
2.3 Discussion
The VPBC methods described have the advantages
of requiring only a few images and little computing
time. However, either they make a certain number of
assumptions or require placing in the scene a calibra-
tion target or a calibrating pattern. Thesemethodsare
unusable if the user has images to be calibrated, but
is unable to access the real 3D scene for including
a calibrating pattern. To remedy this, as in [15–17],
the calibration method described in this paper needs
input data consisting of only one image (a picture)
containing two vanishing points instead of three, as
required by most of the existing VPBC techniques. It
does not need a calibrating pattern or a target.
Recall that [14] provides a means for extracting an
approximated 3D model of the scene background
and for representing the foregroundobjects with bill-
boards from a single image. By means of this, we can
make a tour into the picture without the need of cam-
era calibration or actual 3D reconstruction. With this
technique, the set of allowed trajectories is limited.