DEVELOPMENT OF A SIMULATOR FOR RELIABLE AND ACCURATE 3D
RECONSTRUCTION FROM A SINGLE VIEW
: a . . : b :
D.Aguilera, “ , M. A. Claro-Irisarri *, J. Gomez Lahoz *, J. Finat "^, and M.Gonzalo-Tasis 8
*E. Politecnica Superior Avila, Univ. of Salamanca, Spain,(daguilera, fotod)@usal.es
"MoBiVA Group, Lab.2.2, Edificio I-D, Univ. of Valladolid, Spain ,jfinat@agt.uva.es,
marga@infor.uva.es
Commission V, WG 2
KEY WORDS: Photogrammetry, Architecture, Simulation, Reconstruction
ABSTRACT
Reconstruction from a single view of architectural scenes is possible thanks to the automatic identification of structural perspective
elements in the view. Reliable and accurate reconstruction depends strongly on a robust estimation of vanishing points. In this paper
we compare different approaches for estimation of vanishing points, and we justify our choice of Danish estimator. The intersection
of pairs of pencils of perspective lines through vanishing points generates maps of quadrilaterals in the image which provide
automatic grouping criteria for interpreting the scene. Quadrilaterals are not enough for an accurate reconstruction. Triplets of
pencils of perspective lines provide a wire-framed structure of traditional architectonic scenes used for visualization tasks by
constructing trapezoidal and cuboid maps superimposed to views with a low computational cost. By identifying multiple junctions in
images we obtain the relative orientation between adjacent planes containing bundles of perspective lines. The introduction of
multivector representation simplifies the visualization and management of 3D information arising from the lifting of adjacent
quadrilaterals to 3D cuboids for visualization tasks.
1. INTRODUCTION
Increasing needs for the integration of Architectural
Photogrammetry, Motion Planning, Simulation and Augmented
Reality Visualization require flexible and friendly interfaces for
3D reconstruction, which can be easily updated by minimizing
the human intervention. A robust estimation of vanishing points
is crucial for an accurate and reliable 3D reconstruction of
architectural scenes. Vanishing points arise from the
intersection of a bundle of parallel lines in the scene, which
become perspective lines in its representation. Thus, vanishing
points are traditionally estimated from intersections of putative
perspective lines or by minimizing the distance to projections
of parallel lines. Both procedures are not robust ones, and errors
are unacceptable for photogrammetric applications. Then, we
apply a robust methodology support by the area's triangle
minimization and robust estimators in order to obtain more
accurate results.
Different interesting approaches are developed in [Wil02],
[Heu98] and [Lut94], which are based on the simultancous
estimation of the best projection matrix for a parallelepiped,
parallel and colinearity constraints application and hierarchical
Hough transform respectively.
A robust and accurate identification of vanishing points allows
to superimpose in an automatic way a perspective model on the
view by taking in account the intersections of bundles of
coplanar perspective lines. The line through two vanishing
points is a vanishing line. In the same way, a vanishing plane
contains 3 non-aligned vanishing points. We shall label as
perspective pencils, perspective nets and perspective webs to
the one, two and three-parameter families of bundles of
perspective lines. Perspective lines (resp. planes) pass through a
vanishing point (resp. lines). A perspective line (resp. plane) is
not in general a vanishing line (resp. plane). An automatic
realistic rendering of perspective planes must be take in account
their mutual intersections, the information about visibility in the
scene, and to delete volumetric regions which can occlude
visible regions of background w.r.t. the chosen cuboids in the
foreground. Thus, to avoid the manual intervention it is
necessary to maintain simultaneous information about
perspective lines and perspective planes, and their incidence
relations. The incidence information is performed in terms of
doubly connected lists linked to "oriented projective flags".
Oriented projective flags are given as pairs of an incidence
variety {(l;, m;) | I; © m;} and contain the information relative to
oriented projective lines I; (pencils of perspective lines) and
projective planes m; (nets of perspective lines). The relative
orientation of the wired structure superimposed to the image
arises from identifying the type of double (L and T-type) and
triple (Y and T-type) junctions linked to the intersections
between perspective planes, and contents retrieval assigned to
each junction type. Junctions can be estimated from a Deriche's
filter or a Harris corner's detector. Advances in automatic
visualization of 3D information and interactive visual
navigation require the identification and management of
cuboids in 3D structures linked to the intersection of pairs of
pencils of perspective planes through vanishing lines. Both of
them require to identify structural constraints as collinearity and
coplanarity, and incidence constraints for automatic grouping of
perspective lines. Often in photogrammetry, perspective lines in
architectural scenes are manually drawn to avoid sensitivity
errors. In our case, vanishing points are estimated by a robust
methodology. From vanishing points we retrace perspective
lines through vanishing points in an automatic way, by taking
in account clouds of mini-segments which are grouped
following some variant of non-linear regression or RANSAC
methods [Tor03]. In this way, we obtain bundles of perspective
lines, by avoiding the lack of robustness of least squares
methods linked to an automatic grouping of mini-segments,
which is a specially cumbersome problem in highly textured
images. There are essentially two approaches to recover and
visualize a 3D structure: features and primitive-based
approaches. In the current work, we have followed a primitive-
Internation
AE EN
based appr
better vectc
of vanish
methodolog
with robust
between vi
quadrilaterz
Planar coll
collineation
from the th
managemer
an initial ci
from pencil
provide the
model. Thi
allows to
coherence
constructiot
motion [Fir
and outdooi
problems
concave/co1
results.
With the g«
robust meth
us to test ai
conditions s
next scheme
| SINGLE
Compu
vanishing
Perspe:
reconstn
M
2.1 Auton
Automatic
performed
decomposit
convex sets
[Ber97] Mi
Minisegmet
multiple jur
the slope ar
the striped
minisegmen
detector. Fu
