Bofang Zhao 
  
extraction. The sizes of buildings vary greatly. Tests have been carried out on a rectangular portion, marked on Figure 
3(a). The length and width of the rectangle also represent the X and Y directions of DSM to be generated. 
A DSM was generated on Socet Set digital photogrammetric workstation, with a ground sample distance (GSD) of 0.5 
meters in both X and Y directions, as is shown in Figure 4. The generated DSM was segmented to derive initial regions 
of interest (ROIs), using the method described in Section 3.1. The height threshold parameter used for segmentation is 3 
metres. The initial ROIs were projected back onto the images via collinear equations. Subsequent processing has been 
performed on both left and right images, but only those results yielded from the left image are presented below. Figure 
5 presents the results of initial DSM segmentation in image space. 
  
Figure 4. DSM 
   
Figure 6. Building Shadow Lines Figure 7. Detected Building Regions 
The initial ROIs were refined with texture information and verified by building shadow lines, which were extracted 
using approaches presented in Section 3.2 and 3.3 respectively. Figure 6 illustrates the extracted building shadow lines. 
Figure 7 shows the final ROIs, from which it can be seen that all buildings in the test area have been successfully 
detected and separated from other objects. Based on the detected building regions, buildings were extracted and 
reconstructed locally through all steps described in Section 4. Figure 8 shows a set of processing results carried out on a 
window of 500 x 500, presenting the original image (a), extracted linear features (b), detected building regions (c) and 
the extracted building boundaries (d) respectively. 
6. CONCLUSIONS AND FUTURE WORK 
In this paper, we have presented a hierarchical framework for automatic building extraction from aerial images. It is 
based on the integration of multiple information, knowledge and algorithms, with close cooperation between 2D and 3D 
information. The whole task is accomplished through global detection to local extraction and modelling. The system 
applies a generic polyhedral model, and is hence potentially able to handle buildings of different shapes and structures. 
  
1030 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.