Window Detection in Facades for Aerial Texture Files of 3D CityGML Models

Franziska Lippoldt; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 11-19

Abstract


The author inspects the optimal way to extract geometric facade features of windows from aerial texture files of CityGML models. The following method can be integrated and used for aerial texture modifications or 3D modeling details of 3D CityGML models. The author uses the Mask R-CNN with different configurations and backbone graphs to be tested on two data sets. As to improve the scores on the data sets, two traditional solutions to adjust the results are used: The author tests to integrate the more traditional approach of dbscan clustering to correct the results. Further the author also uses the texture coordinates available from the 3D CityGML file to correct our predictions. As those 3D model textures origin from aerial photos, but are essentially smaller crops of a bigger image, facing typical challenges associated with low-level vision problems and bad image resolution and quality. This application can detect windows and facades from the Berlin CityGML model, extract the windows and doors and adjust the 3D model to integrate those. In addition, it is possible to replace the original windows and doors and insert black counterparts or standard models. The latter procedure will play a crucial role in privacy, as those elements might reveal private objects or persons next to the windows and can be automatically replaced.

Related Material


[pdf]
[bibtex]
@InProceedings{Lippoldt_2019_CVPR_Workshops,
author = {Lippoldt, Franziska},
title = {Window Detection in Facades for Aerial Texture Files of 3D CityGML Models},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}