642 
Rottensteiner, Grussenmeyer, Geneva 
mathematical model of camera distortion can be changed, cameras can be re-assigned to photographs, and adjustment can be 
repeated. 
Creating the geometrical model: Using the interactive measurement tools of ORPHEUS, a wire frame model of the object can be 
created by measuring the relevant object vertices and defining lines between these vertices. In the current version, it is necessary to 
define these lines in a way to create triangles so that several conditions are fulfilled: (1) each part of the object has to be covered 
exactly by one triangle, (2) these triangles must not intersect, (3) there must not be three triangles sharing a common edge, and (4), 
for each triangle there has to be at least one image where all its vertices have been measured. As soon as all object vertices have been 
measured in the images, their object co-ordinates are determined by spatial intersection, and a boundary representation (B-rep) is 
created by finding closed triangular loops in the wire frame model. Contradictions to the rules described above are reported, and the 
triangles involved are discarded. Neighbouring co-planar triangles are merged to closed polygons. The resulting B-rep model is 
exported in the VRML format (Carey and Bell 1997). Errors can be corrected according to a visual inspection. GESTALTs can be 
used for determining points only visible in one image. The creation of the geometrical model is terminated by a final overall 
adjustment of all available observations. 
Texture mapping: As soon as a consistent geometrical model has been generated, texture from the digital images is projected to the 
faces of the B-rep. In ORPHEUS, the “optimal” photo is being searched for, i.e., the photo with a viewing direction closest to the 
normal vector of the face (Dorffner and Forkert 1998). This selection can be overridden by the user after a visual inspection of the 
photo model. Eventually, contrast enhancement tools can be applied to improve the visual appearance of the photo model created in 
the VRML format. 
3. HANDLING THE ZURICH CITY HALL DATA SET USING ORPHEUS 
3d photo models of the Zurich city hall were created from both the Fuji and the Olympus data sets using ORPHEUS. From the 
Olympus data set, all available photos were used. With respect to the Fuji data set, images 2 and 3 were left aside. The camera 
distortion coefficients were estimated from the distortion tables given in (Streilein et al. 1999). 
3.1 Orientation of the images 
Two variants of the block were computed for each of the two data sets, both making use of the control points for the determination of 
the geodetic datum of the blocks. In the first variant, the parameters of inner orientation and lens distortion presented in (Streilein et 
al. 1999) were used as constants. The second variant included self-calibration of the cameras. With respect to the Olympus data set, a 
third variant was computed to show how the geodetic datum can be determined just from the definition of plumb lines and from 
measured distances. 
3.1.1 Interactive measurement 
The control points and well-distributed tie points were measured in all images in the way described in section 2.2. With respect to the 
tie points, care was taken to select points visible in at least three images. Some of these points were chosen outside the facades of the 
buildings in order to achieve a good point distribution for self-calibration. For the variants with self-calibration, four GESTALTs 
corresponding to the four vertical building edges were introduced, and for each of these GESTALTs, about five points per image 
were measured (figure 1). As each of these points gives four observations (two GESTALT equations because a line is considered to 
be the intersection of two planes, and two image co-ordinates), but only adds three new unknowns to adjustment (its object co 
ordinates), each point thus measured increases redundancy by one. With the Olympus data set, two additional GESTALTs 
corresponding to the gutter edges of the east and west facades were introduced as horizontal straight lines. Besides of the “block 
points” (control and tie points) and the GESTALT points, the modelling points were also measured. Note that these categories of 
points are not distinguished in ORPHEUS in principle; the distinction is just given by the user, e.g. by selecting different a priori 
r.m.s. errors for different point sets, or by applying a certain numbering scheme. GESTALTs were also introduced for modelling in 
order to determine points on single rays (section 3.1.3). 
3.1.2 Photogrammetric adjustment 
All the observations described in section 3.1.1 were simultaneously adjusted using the bundle block solution of ORIENT. For each of 
the variants described above, robust estimation and data snooping were carried out to eliminate gross errors. A variance component 
analysis was applied to improve the stochastic model of adjustment. After changing the stochastic model, error detection had to be 
repeated. 
Geodetic datum: For four of the variants, the geodetic datum was to be derived from the control points which are part of the data set. 
Performing robust estimation, it was found out that there were contradictions between the control points from different facades. It 
became obvious that the control points were not given in a unique co-ordinate system: the facades appear to be shifted with respect to 
each other. Thus, only the points 101-113 from the west facade were introduced as control points, the other points remaining in the 
block as tie points. Under these assumptions, the average vectors of discrepancies are (0.24 / 1.38) [m] for points 201-212, and (-0.82 
/ -1.05) [m] for points 301-323, the exception being point 311 which has obviously false planimetric co-ordinates in the data set. The 
residuals are not exactly constant over a facade, which might be caused by a false interpretation of the control point sketch in the data