Texture generation of building models from aerial photographs

Student: Marcel Heckel
Supervisor: Dipl.-Medien-Inf. Sören König
Responsible Professor: Prof. Dr. Sefan Gumhold

Motivation

Through the Internet, it is possible to get a variety of aerial photographs that can be used for the generation of textures for digital city models. Often such images have not exact geographic information, that would be necessary to map them on the models. The aim of this work is the elaboration and implementation of methods to find a best possible projection of the photographs onto the corresponding 3D model based on a rough initial estimate of camera position and orientation. The result should then be used to find suitable building texture.

Description

This thesis describes how large 3D models of cities can be textured with the help of aerial photographs. First, the projection parameters are determined with the help of point or line correspondences between aerial photographs and model. Subsequently, the visible facets of the model are determined in every aerial photograph. Then for every facet, the aerial photograph that would supply the best texture for the facet is determined. At the end, the texture pieces are stored effectively in a texture atlas.

Results

In the example implementation, a graphical user interface was written that can generate textures for 3D models. It allows the loading of OBJ models and different images. In addition, point-and line correspondences and be easily set using the mouse. The settings can be saved in project files. Textures and a texture atlas can be generated from the correspondences. Together with the texture atlas, the model with the calculated texture coordinates can be saved in a separate file.

GUI - point correspondences	GUI - line correspondences
GUI -			GUI - line detector
Berlin Alexanderplatz (1)	Berlin Alexanderplatz (2)
Berlin Alexanderplatz (3)	TU Dresden (1)
TU Dresden (2)	TU Dresden (3)

Future work

A big problem is inaccurate data. First, the correspondences could be never determined exactly. This can be compensated by a large number of correspondences. On the other hand, the 3D models are often faulty.

Furthermore, it is ugly, that the correspondence must be created manually. There are procedures that try to recognized edges in the picture and correlate them with edges in the model based on a good initial estimate of the camera parameters.

In the photos, there is usually a different lighting, different as they are times of the day were recorded. This results in different color different intensities and bright lights (shadow).

Download

Unfortunately, the sample data is not available for download for copyright reasons.

• Thesis(.pdf - 8.64 MB)
• Slides(.pdf - 3.00 MB)
• Implementation (.zip - 5.93 MB)
• Video: Demonstration of the application (.mpg - 37.00 MB)