Performance analysis of object space matching for building extraction using several images

In previous work we presented a system, MULTIVIEW, used to generate 3D building hypotheses starting from sparse features, hypothesized building corners, extracted from multiple views of the scene. This technique results on knowledge about the imaging geometry and acquisition parameters to provide rigorous geometric constraints for the matching process. The effectiveness of this approach was demonstrated using complex aerial imagery taken with highly oblique views containing buildings with flat or peaked roofs. The MULTIVIEW system uses multiple views based on the successive incorporation of new image data into an existing partial solution. This approach allows the generation of data not seen in the initial images and increases the 3D positioning accuracy of derived object models by simultaneous solution of the collinearity equations. In this paper we provide a detailed performance analysis on the 3D buildings constructed by MULTIVIEW. We evaluate these results with respect to several issues including: (1) Metric accuracy of building recovery. (2) The ability to improve detection and delineation as the number of views are increased. (3) The effect of image processing order on building detection and delineation. The implications of incremental construction of detailed 3D structures are examined with respect to manually derived ground truth data.