next up previous contents
Next: Conclusion Up: Augmenting video footage Previous: Augmenting video footage   Contents


A first set of experiments was carried out on video sequences of the Béguinage in Leuven (the same as in Figure 9.7). The sequence was recorded with a digital camcorder in progressive-scan mode to avoid interlacing problems. Once the structure and motion has been computed, the next step consists of positioning the virtual objects with respect to the real scene. This process is illustrated in Figure 8.15. The virtual objects are positioned within the computed 3D structure. To allow a precise positioning, feedback is immediately given by rendering the virtual object in some selected key-frames.
Figure 8.15: Béguinage sequence: positioning of virtual object (top), frames of video augmented with cube (bottom).
...=5.5cm}\hspace{1mm}\psfig{figure=mod/AR/,width=5.5cm}} \end{figure}
After satisfactory placement of each single virtual object the computed camera corresponding to each image is used to render the virtual objects on top of the video. Anti-aliasing can be obtained by merging multiple views of the virtual objects obtained with a small offset on the principal point. Some frames of the Béguinage video sequence augmented with a cube are also shown in Figure 8.15.

Another example was recorded at Sagalassos in Turkey, where the footage of the ruins of an ancient fountain was taken. The fountain video sequence consists of 250 frames. A large part of the original monument is missing. Based on results of archaeological excavations and architectural studies, it was possible to generate a virtual copy of the missing part. Using the proposed approach the virtual reconstruction could be placed back on the remains of the original monument, at least in the recorded video sequence. This material is of great interest to the archaeologists, not only for education and dissemination, but also for fund raising to achieve a real restoration of the fountain. The top part of Figure 8.16 shows a top view of the recovered structure before and after bundle-adjustment. Besides the larger reconstruction error it can also be noticed that the non-refined structure is slightly bent. This effect mostly comes from not taking the radial distortion into account in the initial structure recovery. Therefore, a bundle adjustment that did not model radial distortion would not yield satisfying results. In the rest of Figure 8.16 some frames of the augmented video are shown.

Figure 8.16: Fusion of real and virtual fountain parts. Top: structure-and-motion recovery before and after bundle adjustment. Bottom: 6 of the 250 frames of the fused video sequence

next up previous contents
Next: Conclusion Up: Augmenting video footage Previous: Augmenting video footage   Contents
Marc Pollefeys 2002-11-22