3D Reconstruction of Buildings From Images with Automatic Facade Refinement

Studenteropgave: Speciale (inkl. HD afgangsprojekt)

  • Christian Lindequist Larsen
4. semester, Vision, Grafik og Interaktive Systemer, Kandidat (Kandidatuddannelse)
This project deals with the problem of reconstructing 3D models of real world buildings from images. Potentially real estate marketing can be improved by providing interactive visualizations, e.g. on websites, of properties for sale. For this to be a viable option, however, simpler methods for 3D reconstruction are needed. In particular existing user assisted methods can be improved by automating the process of adding facade details to the reconstructed models.

In this project a proof of concept system covering the whole reconstruction process has been developed. Structure and motion is recovered from an unordered set of images of the building to reconstruct, and this is followed by user assisted reconstruction of a coarse textured 3D model. The primary contribution in the project is the development of a novel method for automatic facade reconstruction, which when applied to the coarse model automatically adds facade details such as recessed windows and doors. The proposed method is based on analyzing the appearance of the facade, and this is achieved using methods for image processing and pattern classification.

The results obtained from using the developed system for reconstructing several refined 3D models of buildings from images show that the proposed reconstruction method is suitable for this purpose. The developed method for automatic facade reconstruction on average correctly detected and reconstructed 89% of recessed windows for the tested buildings. Some work remain for automatic facade reconstruction to be fully automatic, e.g. the depth of recessed regions is specified by the user, but in general it is concluded that the proposed method leads to an improvement compared to existing user assisted reconstruction methods.
Udgivelsesdato3 jun. 2010
Antal sider125
Udgivende institutionAalborg University
ID: 32366959