RUB arrow Bauwesen arrow Informatik arrow


In recent years, automated recognition and 3D reconstruction of existing buildings got more and more into the focus of several fields of research. The spectrum of applications increases steadily demanding diverse levels of detail (LoD) of reconstructed models. For navigation tasks usually coarse block models of buildings are already sufficient. However, for localization in urban areas for, e.g, tourist guidance systems, the recognition of meaningful facade elements is required. Virtual reality applications as well as game development require models of existing buildings with a high LoD to generate realistic virtual environments inducing an immersive experience. In civil engineering, models with different LoD are used for various issues. The possibilities of application range from urban planning which requires only low LoD up to simulation tasks often demanding high detailed models. Especially in the subproject D3 of the SFB 837 high LoD models are mandatory to simulate existing buildings' stiffness with regard to a precise risk assessment of tunneling induced damages.

Background and Problem Statement

Windows are the most prominent facade elements and additionally provide further information. From them the number of stories or even a building's stiffness can be derived. Furthermore, a building can be identified by the alignment of its windows. Thus, windows are highly relevant for recognition and 3D reconstruction tasks as they enrich generated models with a variety of information.
Although some approaches already address the detection of windows in facade images it has not sufficiently been solved yet and still remains to be a challenging task. Most previous approaches detect windows by alignments and patterns which they usually form on facades. However, this may fail for following reasons:

  • The facade exhibits only few windows so that there is no detectable pattern
  • The facade is too complex so that patterns get broken


To avoid aforementioned issues, we develop a detection system which identifies windows by their inherent characteristics represented by image features. Our System is structured as follows:

  • Preprocessing
    • rectification of facade images
  • Detection
    • scanning the image
    • classifying image patches
    • merge detections



Due to their diverse sizes and shapes windows innately possess a high intra-class variability. To reduce variability caused by external factors, e.g, the camera angle, the facade images are normalized before applying our detection system. For that the images are rectified to unify the windows appearance and normalized regarding their brightness.



For detection we apply a naive sliding window detector which shifts an rectangular area across the image in multiple scales and aspect ratios. At each position the underlying image patch is passed to a cascaded classifier consisting of thresholded Haar-like features. The classifier evaluates the particular image patch by applying the features in a cascading manner and returns the result to the detector. As the cascaded classifier is invariant to small translations, detections may also occur around windows. Hence, overlapping detections are subsequently merged.


Evaluation and Results

For training the classifier of our detection system we use randomly chosen window samples of the CMP Facade Database which consists of 378 rectified facade images from various countries. We evaluate our system on the Ecole Central Paris Facades Database with rectified facade images from France, Spain, Greece, Hungary, Romania, and USA. We apply our system to randomly chosen images from each country. The system yields a detection rate of about 85% on average while exhibiting 2% of misclassifications.

res1 res2

Project data

Title: Window Detection in Facade Images

Type: Internal Project

Researcher: Marcel Neuhausen, M.Sc.

Chair of Computing in Engineering
Bldg. IC, Room 6-79/81  
Universitätsstraße 150 
44780 Bochum