Nevatia - 8
Figure 7. Detected roofs from multiple views
simplicity. Simultaneous matching, however, ensures global consistency. In the case of roofs of
buildings, the matching of lower level features is done pairwise, in an order-independent way,
b) Scene Registration and Model Validation :
In this task, we wish to register two or more images to each other, or match a model of the
scene with the image which may be a preliminary step in validating whether the model is still
accurate. In these tasks, we can resort to global matching of features if the fixed objects, such as
roads and buildings, are the dominant features. The global matching also allows us to use lower
level features, such as line segments, for matching. Multiple matches may still be found but they
representations of the man-made objects in the scene) to a new image [1]. Besides scene
registration, this system can also verify whether the visible structures in the model are still present
in the scene. Fig. 8(a) shows a projection of the 3D model of the scene in Figure 1, and Figure
8(b) shows the model registered with an image.
c) Indoor Object Recognition :
Most objects in indoor scenes are either mobile or moveable. Thus, we can not rely on a
global match but need to recognize individual objects separately. It may still be possible to use
some context such as the observation that objects like a telephone or a computer monitor are
likely to be found on top of a desk.
When structures to be matched are local, it is more important to use higher level
while the hypothesis of the higher level structures, such as U-s and roofs is done simultaneously.
should be small in number and higher level descriptions can be used to discriminate among them.
We have applied such a method in matching the site model of a scene (consisting of wire-frame