Image-Based 3D Acquisition Tool for Architectural Conservation 
303 
In addition to the epipolar geometry other constraints like preserving the order of neighbouring pixels, bi-directional uniqueness of 
the match, and detection of occlusions can be exploited. These constraints are used to guide the correspondence towards the most 
probable scan-line match using a dynamic programming scheme (Cox et al. 1996). The matcher searches at each pixel in one image 
for maximum normalized cross correlation in the other image by shifting a small measurement window along the corresponding scan 
line. Matching ambiguities are resolved by exploiting the ordering constraint in the dynamic programming approach (Koch, 1996). 
The algorithm was further adapted to employ extended neighbourhood relationships and a pyramidal estimation scheme to reliably 
deal with very large disparity ranges of over 50% of image size (Falkenhagen, 1997). Using the common reference points seen in the 
two images, the disparity search range is limited to the proximity of the scene. This results in much lower computation times and, in 
addition, improves the robustness by reducing the possibility for mismatches. 
From the disparity estimates the 3D points can be reconstructed through triangulation and the depth can be computed as the distance 
between the point and the center of projection of the reference view. These depth values are stored in a depth map (an image where 
at each pixel the depth value is stored. On the left of Fig. 5 a depth map obtained from the image pair of Fig. 4 is shown. In this case 
darker means closer. 
In the context of heritage conservation it is important to also have an estimate of the error that is present on the measurements. 
Therefore, error estimation was also implemented. Assuming the dense stereo matching to be accurate within e pixels, one can 
calculate the uncertainty margin on the depth estimate by calculating a new depth for a corresponding point displaced by a distance e. 
For camera viewpoints that are located close to each other, a small difference in the image yields a large error on the estimate. On 
the other hand, with cameras more orthogonal to each other the error is much more contained. This error estimate also depends on 
the distance of the scene to the cameras. On the right side of Fig. 5 a map of error estimates is shown. One can observe that indeed 
both maps look very similar (up to scale). This is because points further away have more uncertainty than points closer to the 
cameras. The average error estimate is around 3cm in this case. 
Fig. 5: Depth map (left) and map containing error estimate (right) 
4. 3D MODELLING AND ANALYSIS 
In the previous sections both the camera calibration and dense depth maps where computed. This yields all the necessary 
information to build 3D models. On the left of Fig. 6 a 3D reconstruction consisting of the reprojection of a sub-sampled depth image 
is shown. In the system developed in the context of our project, this data is then passed on to another tool developed during this 
project (Lauwers and Li 2001) for integration of the results of different depth maps into a single 3D model. On the right side of Fig. 
6 a textured 3D model is shown that was obtained by combining several depth maps and using the original photographs as texture. 
In the context of conservation it is important that the computed 3D model can also be used to obtain measurements from it. As an 
example Fig.7 shows a number of isolines automatically computed by the system. These can for example be exported to a CAD 
package to derive plans from it. More details on this and other applications can be found in (Nuyts et al. 2001) where the third tool 
of the system is described. 
5. EXAMPLES 
In this section some additional examples are shown. On the left of Fig.8 an image of the reconstruction of a painted vault is shown. 
The vault is located in the Arenberg castle in Leuven. On this reconstruction isolines are also indicated. In the middle of Fig.8 a part 
of the inner court of the Arenberg castle is shown. On the right of Fig.8 a 3D model of ancient tower in St-Truiden can be seen.