3D Modeling from Photographs
Guenter Pomaska, Bielefeld University of Applied Sciences
Gamze Saygi, Izmir Institute of Technology
Short descripton of an architectural documentation project
applying simple photogrammetric methods for image rectification and 3D
modeling.
A cooperation between Bielefeld University of Applied
Sciences, Izmir Institute of Technology and Aalborg University, funded by
European Erasmus programme 2007.
Digital Image Aqcuisition
Common digital cameras are available as digital compact cameras (DCC)or single lens reflection cameras (SLR). While a former 35 mm analog camera was equipped with a film format of 24mm x 36mm, we have today various senor formats, due to the type of the sensor. Sensor types are announced as fractions like 1/2.7", 1/1.8" or 2/3". This type designation harks back to TV camera specifications. The table below lists the dimensions of a DCC and SLR camera. Both sensors came with a high resolution of about 7MPx. Depending upon the smaller CCD area, the pixel size differs. Smaller pixels effect in a lower image quality.
| Type | Aspect Ratio | Diagonal | Width | Height | Resolution | Pixel Size |
| 1/1.8 | 1.3 | 7,18 | 5.76 | 4.29 | 3072x2304 | 0.0018 |
| 1/2.3 | 1.5 | 28,66 | 24.06 | 15.58 | 3008x1960 | 0.0080 |
The sensor areas compared to an analog film surface are shown in the image below.
Image Refinement
We don't focus here on the different kind of sensor types like the Foveon sensor, Fuji Super CCD, CCD or CMOS sensors. Pixel quality is effected by geometric accuracy, color accuracy, dynamic range, noise and artefacts. According to our application, it is important to reduce the impact of lens distortion. Lens distortion causes a square in an image distorted in barrel or pincushion form. Barrel distortion is associated with wide angle lenses and pincushion distortion is most detectable at the end of zoom tele lenses.
There are several possibilities to determine the distortion parameter, as well as software is available for image refinement. Here a test shape is used together with the freeware ShiftN. ShiftN displays the original image together with the EXIF data and can reduce tilt automatically. Estimating the lens distortion value for the radial symmetric amount has to be carried out and checked for every focal length used. Image refinement for multiple pictures can be done in a batch process. The GIF animation, shown here, switches between a corrected and an uncorrected image and highlights the impact of distortion for a wide angle lens.
After the lens distortion correction, which had be done with the
complete image, further image refinement regarding radiometric improvement and
cutting out the areas of interest can be carried out before
rectification.
See the impact of lens
distortion - click the image
Single Image Rectification
Orthophoto production is a process to convert perspectives into orthographic views. It is also known as differential or parametric rectification. This requires a digital object model and camera parameters.
Single image rectification without the use of camera parameter can be carried out by projective transfomation with four control points or vanishing point algorithms. In both cases an object plane is estimated. The latter makes use of two pairs of parallel lines, which can most often be find in architectural objects. The required scale may come from a digital map. Rectification delivers a necessary height of the object and a unique object resolution over the complete project. For education purposes the software VeCAD from Vassilios Tsioukas, Democritos University of Thrace, is available by kindly permission of the author.
Comparable to the photogrammetric procedure is texture mapping
known from computer graphis. A 2D texture corresponds to a 3D object via
according points. But consider we ´don't have a 3D model yet. It has to
be constructed with the height information from the rectified
images.
Mastering Google SketchUp
Before you use an engine, you should know about it. Applying
SketchUp is not an exception from that rule. One has to organize his file into
components, groups and layers and must have knowledge of the inference engine.
SU tries to stick elements together. Grouping protects one geometry from
others. Google SU is a polygon surface modeler. Geometry consists of edges and
faces. Deleting faces does not delete edges, deleting an edge, deletes the
complete face. Despite the ease of use of SU, it is strictly recommended to be
familiar with the tools before modeling from meaurements or from images with
this powerful tool. The funcionality is great and the export functions allow
amongst others conversion to X3D or Google Earth.
The Workflow
Application Ritterstrasse, Minden and Kennedy Plads, Aalborg
Ritterstrasse is part of a test area working on a digital city model. Given is the layout of the houses from a digital map. From rectification we get the heights of the fassades and can copy the base line in the blue direction (SketchUp jargon) about this value. Conneting start- and endpoints with lines generates faces for texture mapping. Further modeling of gables is possible from here on. More geometry is added to the facades by making use of the images.
Another example are the building at Kennedy Plads in Aalborg,Denmark. This sample was worked out during lectures at the Institut for Development and Planning to demonstrate the ease of use the process, documented in the workflow chart above.
See more higher resolution images in the gallery!
