In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
99 
Figure 4. Example of a “missed” wall, (a) 10 rays per camera: 
the current wall is “missed” by several cameras (red 
dots); (b) 50 rays per camera: the current wall is 
seen by all the cameras (black dots) 
(a) 3D view 
(b) Cameras seeing the (c) Cameras selected with 
red wall the 2D approach 
Figure 5. Example of incomplete camera selection 
4. 3D Z-BUFFERING 
4.1 Principle 
The second approach is 3D-based and relies on a z-buffer 
technique. Each camera is analysed in turn. A set of candidate 
walls is first associated to the current camera as in described in 
section 3, using distance, half-plane and backface culling 
criteria. The camera is then associated to a label image 
identifying the walls seen by the camera, and a depth image 
indicating the distance from the camera centre to the walls. 
Finally, after all the cameras have been processed, each wall 
can be associated to the list of cameras that can view it. 
4.2 Test results 
In order to reduce computing time, the distance and label 
images are sub-sampled at coarser resolutions. The tests were 
performed at a sampling resolution of 5, 10 and 20 pixels. 
They are shown in Table 2. It is important to note that the 
algorithm was not optimised and the graphical card not used. 
An example of depth image is shown in Figure 6. 
Z-buffer 
resolution 
Image size 
(hxw) 
Total # of 
visible 
walls 
Avg # of cam. 
per wall 
Computing 
time 
5 pixels 
216x384 
2310 
(20.2%) 
4.40 
61min58s 
10 pixels 
108x192 
2249 
(19.7%) 
4.41 
52min36s 
20 pixels 
54x96 
2186 
(19.2%) 
4.35 
43min54s 
Table 2. Results of 3D ray tracing 
Figure 6. Example of a depth image 
4.3 Discussion 
Using this approach, 50% more walls can be textured. In 
particular, all facades located behind other buildings can now 
be textured, whereas they were discarded with the 2D approach. 
In the example of Figure lOd, the circled area shows an 
example of a high building visible from the current camera but 
only selected with the 3D approach. As the measures are very 
dense, even small walls, walls distant from the path or wall 
aligned with the path can theoretically be textured. 
In return, many selected walls are only partly visible, and 
would actually have a very poor texture quality. It is important 
to introduce a contribution culling technique, in order to discard 
wall images inappropriate for texture mapping. In the current 
implementation, another drawback of the method is the 
computing time. Using a hardware implementation directly into 
the Graphical Processing Unit of the graphic card should solve 
this problem. A hierarchical z-buffer technique could also be 
investigated (Greene, 1993). Finally, the selection process must 
be entirely completed before being able to further process a 
façade, which might not be compatible with a large-scale 
production process. 
5. 3D RAY TRACING 
5.1 Principle 
The last approach combines the main advantages of the two 
previous ones: speed and use of 3D information. It is a 3D 
extension of the 2D approach based on ray tracing. However, 
the analysis is performed wall-by-wall rather than camera-by 
camera. Given a wall, a set of candidate cameras is selected 
using a method similar to the one described in section 3: