es
of
es
ne
ns
ts
ic
Dowman, Ian
e Laser range scanning:
Laser rangers are used to reconstruct the scene structure by direct measurement of depth to various points in the
scene. Once the geometry of the scene structure is built, the images can be mapped to their real positions to form
3D realistic models. While existing versions of these sensors are generally slow, cumbersome and expensive, active
development of this technology is making it of practical use for more and more applications (Debevec, 1999).
e [mage-based rendering techniques:
In a simple case images are projected to the body of a cylinder having a fixed radius from the camera projection centre.
A Quick Time Virtual Reality model (QTVR) is an example of such which is usually built by mosaicing photographs
taken on a virtual sphere around the camera. In more sophisticated situations, the model is reconstructed using a
set of images and their corresponding depth maps. In other words, using the images the structure of the scene is
determined and the images are projected to their determined 3D positions. Some of these techniques employ the
computer vision technique of computational stereopsis to automatically construct the structure of the scene from
the multiple photographs available. As a consequence, these systems are only as strong as the underlying stereo-
correspondence algorithms. This has caused problems because the stereo algorithms have a number of significant
weaknesses; in particular the photographs need to have similar viewpoints in order for reliable results to be obtained.
These techniques are still subject to research and development and are mainly suitable for small projects.
In general both of the above techniques have the advantage of being able to produce truly 3D and realistic models of
buildings. However, they form the model using a huge number of unstructured points. As a result the topology of the
scene is either completely lost or is local to the location of the 3D points. In contrast, the system presented in this article,
is geometry-based and uses a geometric model to define the position of textures in space and to process the textures of
building facades. A Numerical Frame of Reference (NFR), a Texture Image Database (TID), as the source for textures,
and an Automatic Texture Processing Tool (ATPT) are used to process the textures. These components are described
and examined briefly in the following sections. Further discussions and analysis can be found in Varshosaz (1999) and
http://www.ge.ucl.ac.uk/staff/ucfsmva.html.
2 THE NUMERICAL FRAME OF REFERENCE
The NFR is a model that shows the geometric structure of buildings which can be rendered using photo-textures to produce
realistic models. Most of the work carried out in the reconstruction of 3D models of buildings, has concentrated on the
development of techniques for production of NFR models. A NFR is primarily used for making crude 3D measurements
and calculating volumes. It carries enough realism to satisfy needs of applications such as small scale town planning or
environmental monitoring which require only a simple description of buildings. However, as far as the texture processing
is concerned, it is used to define the location, dimensions, and direction of the textures in space. The coordinates of
building corners forming the NFR can be derived from various sources and methods. However, as long as this information
is available and is of sufficient accuracy the way it is provided is not important.
3 THE TEXTURE IMAGE DATABASE
The TID is an indexed archive of spatially registered CCD images, acquired as described in the next section, and includes
the following data:
e Images as captured by the method outlined later;
Reference file: this file determines a number of elements for each image and enables automatic search for images
that cover a particular section of a building face;
Camera file(s): a camera file includes all details about the internal and external geometry of the CCD camera used
for image acquisition;
Station file: this file includes the exterior orientation parameters of the stations at which the images are taken.
To form a TID the images are captured and the reference file is created. To obtain the station and camera files, the images
need to be registered.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 181