CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
100 
The camera is located within a given distance from 
the wall (distance measured at closest point). 
The camera center point is located in the half-space in 
front of the wall. 
The camera is facing the wall plane. 
In order to reduce computing time and improve texture quality, 
an additional criterion has been introduced: cameras that are 
almost aligned with the wall plane are discarded. A maximum 
threshold on the angle defined by the wall plane and the camera 
directions is introduced (see Figure 7). This filtering step is an 
extension of the backface culling criterion. 
Figure 7. Angle criterion for the pre-selection of candidate 
cameras: the cameras with their direction vector in 
the red angular area are discarded 
For each candidate camera a grid on the camera plane is 
defined. Each grid point defines a 3D ray passing through the 
camera center point. The 3D rays not intersecting the current 
wall are ignored. The remaining 3D rays are tested with respect 
to all the walls compatible with the camera (pre-selection 
method described in section 3): any ray intersecting a wall face 
closer than the current one is discarded. The candidate camera 
is finally selected as viewing the current wall, if at least one of 
the rays has not been discarded. The process is illustrated in 
Figure 8. 
Figure 8. Principle of 3D ray tracing: the rays launched from 
the tested camera are discarded if they do not 
intersect the current wall (see rays on the extreme 
sides) or if they first intersect a closer wall (see rays 
on the left) 
5.2 Test results 
The method was tested with 10x10 and 20x20 rays per camera, 
with and without threshold on the angle during pre-selection. 
The threshold on angle was set to 3# radians when applied. The 
8 
distance threshold was set to 150m (identical to 2D ray tracing). 
Numerical results are shown in Table 3. An example of selected 
walls is illustrated in Figure lOd. 
Method 
Total # of 
visible walls 
Avg # of 
cam. per wall 
Computing 
time 
I Ox 10 rays 
1349(11.8%) 
4.36 
3min51s 
20x20 rays 
1604(14%) 
4.49 
1 lmin45s 
10x10 rays 
amax = 3*n/8 
1032 (9%) 
4.55 
2min49s 
20x20 rays 
amax = 3*11/8 
1213 (10.6%) 
4.56 
8min25s 
Table 3. Results of 3D ray tracing 
5.3 Discussion 
As expected from a 3D-based approach, the walls located at the 
background can be textured if they are high enough. Fewer 
texture images are selected with the 3D ray-tracing approach 
than with the z-buffer approach, but they generally have a 
better quality. It is not surprising as ray tracing is not a dense 
approach and most small wall textures are naturally discarded. 
In the example of Figure lOe, only the relevant facade of the 
high building located at the back of the block was selected as 
visible. Figure 11 shows another example of distant facade that 
can be textured only with a 3D approach. 
The additional pre-selection criterion on angles removes 
sidelong walls, which are usually seen by few cameras. It 
improves the relevance of the selection by discarding walls 
with a poor texture resolution. Using 20x20 rays instead of 
10x10 rays significantly increases the total number of visible 
walls, but further tests are needed in order to find out whether 
these additional walls can be textured with a good enough 
quality. Importantly, as each wall is processed in turn, the 
texturing stage can be performed without requiring the 
complete processing of the path. 
The computing time is intermediate between 2D ray tracing and 
3D z-buffering. In our implementation many calculations are 
redundant. A spatial division of the space could be performed in 
order to make use of object-space coherence and accelerate ray 
tracing (Glassner, 1984; Jevans and Wyvill, 1989). 
6. CONCLUSION AND PERSPECTIVES 
The 2D approach is satisfactory in most cases, and it is fast, 
simple and easy to implement. However any building located 
behind another cannot be textured. 
The 3D approaches provide more exhaustive wall textures, 
including texture images for high building walls located at the 
back of lower buildings. The use of the 3D dimension makes 
the visibility estimation closer to ground truth, and the selection 
process more efficient. Although both ray-tracing and z-buffer 
techniques can be implemented very efficiently, the approach 
based on 3D ray tracing is a good compromise to achieve a 
relevant selection. It also seems important to prefer a wall-by- 
wall analysis, as further texturing stages can then be performed 
without requiring the complete processing of the path. The z- 
buffering technique could be considered if the resulting depth 
image is a valuable source of information in further steps. 
The main constraint for the 3D approaches is obviously the 
availability of a 3D building database. Given a 2D map, the 3D 
information can be derived from a correlation-based or LIDAR 
Digital Elevation Model, or even from the analysis of 
architectural plans or building permits. In our opinion, a coarse 
3D city model is sufficient to significantly improve the 
relevance of the texture selection. 
We are now working on refining the selection with texture 
quality criteria rather than just visibility. The texture quality