At the start of the 3DVIDEO-project, the accuracies of the pointwise measurements were of the order of 1 : 10,000 to 1 : 40,000 depending on the number of cameras in the network. The measuring times considered as real-time operations were within seconds or less than a minute depending on the application requirements. Neither profilewise nor surfacewise measurements were applied in operative applications in Finland at that time.
During the 3DVIDEO-project, we achieved accuracies of the order of 1 : 5,000 to 1 : 15,000 for the profilewise measurements using the developed techniques. The procedure was applied for dynamic processing of both living and recorded image sequences. The established procedure produces a detailed array of samples for a complete analysis of the object geometry. It is accurate of utmost order of 1 : 5,000 relative to the single scene dimensions. The system acts as a movable submodule of a photogrammetric station using its coordinate system as a reference datum. The accuracy of the initial digitizing is of the order of 1 : 1,000 - 1 : 5,000 for the pixels where the disparity has been measured and less for the interpolated ones. The accuracy of the disparities after the full digitizing were of the order of 1 : 5,000 - 1: 10,000 relative to the dimensions of the scene that is visible.
At the start of the 3DVIDEO2-project, the registration problem was internationally recognized as unsolved. In general, our results are as good as any others working in this field.
In the 3DVIDEO2-project, the registration of two disparity maps was tested with synthetic data [ ]. In this case, the feature based method gave more accurate results than the area based one when the modeling of features was successful. On the other hand, the area based method was capable of handling cases with lack of features.
The registration of multiple profile maps was tested with real data
[ ].
In the first experiment, the object consisted of a casing box on a plate
and four views over the scene were acquired from different viewpoints.
The computation time for preprocessing all the four maps without the
interactive step in segmentation was 1 min 40 seconds.
Using the area based matching, we achieved results were the root
mean sum of the squares of weighted distances was 
pixels.
This is of the same order as the measuring accuracy of
the profilometer that is estimated to be about half a pixel since the
laser stripe is seen as fairly broad in the image of the camera.
The computation time for the refinement of the registration
was 7 min 50 seconds for 26 iterations.
In the second experiment, the scene was made up of a scale model of
an urban area and five profile maps over the scene were measured.
The preprocessing of all the five maps without the interactive step
in segmentation required 3 min 50 seconds computation time.
After surface matching, the root mean sum of the squares
of weighted distances (
pixels) was of the same order as the measuring
accuracy of the profilometer. The computation time for the refinement
was 15 minutes for 18 iterations.