We now use the expert mode to point out some interesting concepts connected with the reconstruction algorithm. Although the model of the house looks pretty simple, it gives us the opportunity to underline some crucial points in a more precise way. For example:
Jumping to the next step, we have to deal with the corner extraction process in both images (Figure 4 (b)): the user (in learning mode) finds already the corners precisely highlighted. He/she can then move to the next step. In the expert mode it is possible to perform a corner search on one (or both) images, to test how the parameters influence the results. For example, increasing/decreasing the size of the Search window the user can see how the corner detection is correspondingly affected. The same can be done with the Threshold parameter.
After completing the corner extraction step the user can move to the next one, corner matching (Figure 4 (c)): again in learning mode the user is given an exact matching of the corners. In the expert mode it is possible to run the matching algorithm on the results obtained in the corner extraction step by using the form button Match. The results can be refined by reducing the Comparison area radius, which usually allows for a higher number of matches at the expense of precision.
|
After estimating the fundamental matrix and roto-translation parameters the next step involves the rectification of the two images. By clicking on Rectify the user obtains from the input stereo pairs the rectified ones (Figure 4 (e)).
Rectified images greatly simplify the dense matching step: the latter is the search of a dense set of correspondences in the two rectified images. After that, using the disparity in the correspondences, it is possible to acquire the information about the 3D positions of points and build a model of the scene. In learning-mode the user already finds a suitable set of correspondences to obtain a reasonably faithful model. In the expert mode, additional sets of parameters are available, like Window size, Threshold and Radius of comparison window.
The final step is the triangulation (Figure 4 (f)). The output is the 3D model and already in the learning-mode is presented to the user in three forms:
In the expert mode, the new model is not given, but obtained from the information inherited from the previous steps, after choosing if the final result should be shaded or be applied suitable textures.
Using the bottom frame the user can move across the various steps, which is useful especially for advanced users. In all cases, on line help is available from the top frame. We spend a few more words about the way this becomes important, whatever the user mode is, when combined with the interactivity available in the middle frame.
Indeed, in many of the steps the user employs this frame to modify the results given by the algorithm, even after setting up modified parameters. For example, let us consider again the matching step: let us assume that the user, after fixing the matching parameters, obtains one wrong match. He/she can go on with the reconstruction process and get the final model, which will present some inconsistencies. He/she can go back and eliminate the wrong match pair, and execute the subsequent reconstruction steps to verify how this affects the final results. Then he/she go back again to correct - not remove! - the wrong match and again verify the impact on the final model. This is helpful to get some clues about improvements that could be necessary in the reconstruction algorithm.
![\includegraphics[width=12cm]{session.eps}](img12.gif)