VOLUMETRIC MODEL REFINEMENT BY SHELL CARVING
Y. Kuzu ?, O. Sinram "
? Yildiz Technical University, Department of Geodesy and Photogrammetry Engineering
34349 Besiktas Istanbul, Turkey - kuzu@yildiz.edu.tr
® Technical University of Berlin, Department of Photogrammetry and Cartography
Str. des 17 Juni 135, EB 9, D-10623 Berlin, Germany - sinram@fpk.tu-berlin.de
Commission V, WG V/2
KEY WORDS: Calibration, Close Range, Correlation, Reconstruction, Visualization.
ABSTRACT:
In this paper we present a voxel-based object reconstruction technique to compute photo realistic volume models of real world
objects from multiple color images. The 3D object acquisition is performed using a CCD video camera, where the images of the
object have been captured in a circular camera setup. Before starting the image acquisition, the sensor is calibrated in order to
determine the interior and exterior parameters of the camera. Due to the missing of control points on the object, the images undergo
a process of relative orientation in a free network using manually measured tie points. The approximate model of the object can be
casily acquired by a volume intersection method in a fast and a robust way, which results in the convex hull of the object. The shell
of the object, namely the surface voxels are easily obtained from this model. In order to get into the concavities and refine the model
we introduced the shell carving method. For the visibility computation, we introduce ray tracing and surface normal vector
computation. We use color image matching to search for image correspondences. Furthermore, we make use of the image orientation
data for a knowledge based template grabbing. Voxels which are projected into non-corresponding image points are carved away
from the shell. And the final set of voxels contains sufficient color and texture information to accurately represent the object.
1. INTRODUCTION
Photogrammetry and computer vision are closely related
disciplines which meet on common research areas such as
creating 3D object models. While high metric accuracy is more
important in photogrammetry, computer vision seeks
automation and speed. Therefore it is beneficial to apply
techniques developed in both disciplines for faster and more
accurate results. In this paper issues from both of these
disciplines are referred to compute volumetric models of
objects from its color images.
3D object reconstruction from a series of digital images is an
important problem both in computer vision and
photogrammetry. Optical 3D shape acquisition can be
performed either by scanning the object or by taking its images.
In this paper the shape of the objects is captured by a CCD
camera which is a low-cost alternative to laser scanners.
The shape recovery method described in this paper can be used
in many application areas such as in virtual worlds,
entertainment, cultural heritage preservation, home shopping
and design.
In this paper we use voxels as 3D primitives. Volumetric data
was first introduced in the 70's in medical imaging and is now
commonly used in scientific visualization, computer vision and
graphics. Although voxels consume large amounts of memory
they provide more flexible reconstructions of complex objects.
Therefore voxel-based reconstruction methods have become an
alternative to surface based representations.
In this paper the model is acquired in two steps. First, the
approximate model is acquired by a volume intersection (shape
from silhouettes) algorithm. The intersection of the silhouette
cones from multiple images gives a good estimation of the true
model which is called the object's visual hull (Matusik et al,
2000). This algorithm is popular in computer vision due to its
fast computation and robustness. However the concavities on an
object cannot be recovered with this technique. We refine the
model acquired by volume intersection method by our shell
carving algorithm. The closest related method to our proposed
algorithm is the Voxel coloring algorithm (Seitz and Dyer,
1997). Also see (Kuzu and Sinram, 2002). These algorithms use
color consistency to distinguish surface points from the other
points in the scene. They use the fact that surface points in a
‚scene project into consistent (similar) colors in the input
images. Our algorithm differs from present voxel coloring
algorithms in the way that we compute the visibility
information. The other basic difference is instead of pooling the
pixels of the images a visible voxel projects into, we perform
image matching powered by knowledge-based patch distortion
to lessen deformation effects.
In the next chapter, the image acquisition setup is described.
The image orientation process is also explained. The
reconstruction of the model using shape from silhouette
technique will be described in chapter 3. Chapter 4 introduces
the computation of visibility information. In chapter 5, image
matching is given and in chapter 6 the refinement algorithm is
explained. Chapter 7 finally summarizes this paper.
2. '3D SHAPE ACQUISITION
We have low-cost system requirements, since we are using à
stationary standard CCD video-camera in order to acquire still
images. We use a calibration object to compute the interior
orientation parameters of the camera. The image acquisition is
performed in front of a blue, homogenous background. Multiple
views are captured rotating the object resulting in a circular
camera setup.
Internationa
ERG LOT
2.1 Camer
Prior the im
We calibrate
object havin
points on ea
switch off th
throughout tl
In a second
frame in ord
We perform
which delive
coordinates
introduced a:
In order to «
oriented, the
known. In n
objects, ther:
The images
We used enc
setup to pe
images. We
orientations,
projection ce
determined v
One of the
objects is
recovers the
A silhouette
image segm(
represent an
background
easily define
pixel’s posit
decide if it
image segme
F
To start. witl
only opaque
points. In orc
the cube's v
defines a bac
giving it the
represents e
intersection
multiple ima,
which is ca
Rodehorst, 2t
In Figure 1
acquired, usi
increasing n
approximatio
As shown in
method since
