Corrective Interaction

If the building is still not correctly detected, additional information is needed and one has to go one step backwards in the hierarchy of USC-MABS to extract new features, like edges or corners. Two ways of correcting the first hypothesis are offered: first the user can adjust the roof-parallelogram by dragging sides with the mouse, rotating, translating or changing the angle of the whole model. Changes can only be made within the constraints of the building model, for example opposite sides remain parallel. The extraction of a ground corner or -edge (shadow corner or -edge) will determine the building height. These interactions are similar to a completely manual system.

Second, one can choose to extract edges and corners and associate them to a part of the building model. For example, a roof-side of the building can be specified by an edge extracted in the image. Then this edge is added to the current hypothesis. (see also Example 3) Our systems are implemented to run under the RCDE. This environment allows the use of mouse-sensitive features thus facilitating user selection and manipulation of features.

After each corrective interaction the system forms a new parallelogram-hypothesis, the system looks for new edges, shadow and wall evidence to support the new hypothesis and finally performs a fitting and verification step. These methods are the same as those in the automatic system. This important step of verifying the consistence to the constraints proposed in the USC-MABS can be compared to a fitting process in a computer assisted manual systems, though in our system a fitting is performed after each interaction. Therefore it is possible that, after a manual correction of a roof-boundary, the wrong building height is also corrected automatically.

Click on the boxes below to get example images or MPEG-movies

Unfortunately, there is no movie for Initial Interaction and Select most likely hypothesis!)

Without the fitting step the system would perform like a manual system and at least three interaction steps (two corner adjustments and one correction of the building height) would be necessary for adjusting the shape of one building model. Rotation and translation as parameters of the position add another two steps.

Note, that the manual feature extraction and the following fitting and verification steps can also be applied to buildings that are automatically detected, but partially wrong.