Next: Conclusion Up: Combined metric structure and Previous: Combined metric structure and   Contents

### Example

In this section results of our approach on two real image sequences are shown. The first image sequence was recorded from a corner of our institute. The corner sequence contains 64 images recorded using a Sony TRV900 digital camcorder in progressive scan mode. The images therefore have a resolution of (PAL). Some of the images are shown in Figure 5.7. Note that the images contain quite some radial distortion.

In Figure 5.8 the GRIC values are given for and as well as for and . It can clearly be seen that -besides dealing with additional ambiguities- the triplet based analysis in general provides more discriminant results. It is also interesting to note that triplet 34-35-36 is clearly indicated as containing sufficiently general structure for the triplet-based approach while the pair-based approach marginally prefers to use the plane based model. The USaM approach reconstructs the structure for this triplet (including some points seen in the background of the lower left picture of Figure 5.7) and successfully integrates them with the rest of the recovered structure and motion.

Figure 5.9 shows results for different stages of our approach. At the top-left the recovered metric structure and motion for the two subsequences that contain sufficiently general structure is given (after coupled self-calibration). Then, both structure and motion are extended over the planar parts. This can be seen in the middle-left part of the figure. At the bottom-left the complete structure and motion for the whole sequence is shown after bundle adjustment. On the right side of the figure orthogonal top and front views are shown.

The second sequence consists of 150 images of an old farmhouse. It was recorded with the same camera as the first sequence. In Figure 5.10 the GRIC values are plotted and for some of them the corresponding images are shown. As can be seen the approach successfully discriminates between the planar parts and the others.

In Figure 5.11 the computed structure and motion is shown. In Figure 5.12 some views of a dense textured 3D model are shown. This model was obtained by computing some depth maps using a stereo algorithm and the obtained metric structure and motion. Note that the whole approach from image sequence to complete 3D model is fully automatic.

Next: Conclusion Up: Combined metric structure and Previous: Combined metric structure and   Contents
Marc Pollefeys 2002-11-22