Deep Learning Training Data for 3D Building Reconstruction

In recent years, machine learning methods have gained in importance through the latest development of artificial intelligence and computer hardware. Particularly approaches based on deep learning have shown that they are able to provide state-of-the-art results for various tasks. However, the direct application of deep learning methods to improve the results of 3D building reconstruction is often not possible due, for example, to the lack of suitable training data. By evaluating a large number of publicly available 3D point cloud training data sets for machine learning, it became clear that a good basis for traditional classification tasks is already available but that the number of classes is still quite limited. Particularly for buildings, which play for most applications in urban areas an essential role, we discovered a shortage in the set of available data sets. According to our knowledge, there is currently no 3D point cloud training data set publicly available that provides distinct classes for buildings. Many applications, however, require a fine subdivision of the building class to distinguish, for example, between different roof types or to recognize certain roof structures.

In order to close this crucial gap, we present RoofN3D which provides new 3D point cloud training datasets for free usage. Further details about the provided data are described in (Wichmann et al., 2018). The provided data can be used to train machine learning models for different tasks in the context of 3D building reconstruction. It can be used, for example, to train semantic segmentation networks or to learn the structure of buildings and the geometric model construction.

Note: RoofN3D is always in progress; the provided datasets will be improved, updated, and extended in the future.


Complete RoofN3D database

Each table is stored as a csv file following our database schema.

Organized data

Each file represents a building and it is organized to be used for training machine learning models.

Source data

    Building points (zipped wkt) [Coming soon]
    Building outlines (zipped wkt) [Coming soon]


    Surface growing (zipped wkt) [Coming soon]
    Sub-surface growing (zipped wkt) [Coming soon]

3D building representation

    Half-space (zipped txt) [Coming soon]
    Boundary (zipped wkt) [Coming soon]
Here should be the image of ED

The architecture of RoofN3D and the provided information about a building


We are grateful to the city of New York for providing building footprints through the New York City Open Data Portal and to the U.S. Geological Survey (USGS) for providing LiDAR point clouds of New York.


  • Agoub, Amgad - Technische Universität Berlin, Germany
  • Kada, Martin - Technische Universität Berlin, Germany
  • Schmidt, Valentina - Technische Universität Berlin, Germany
  • Wichmann, Andreas - Technische Universität Berlin, Germany


Wichmann, A., Agoub, A., Kada, M., 2018. RoofN3D: Deep Learning Training Data for 3D Building Reconstruction. In: The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLII-2, pp. 1191-1198.


Terms of Use and Disclaimer

The RoofN3D training datasets available on this page are provided for free usage by the chair Methods of Geoinformation Science of Technische Universität Berlin. No warranties or guarantee on the correctness, completeness, or appropriateness for any specific use is given. In no case, the chair Methods of Geoinformation Science can be held liable for any directly or indirectly damages caused by the use of these datasets. The training datasets are exclusively based on datasets from the New York City Open Data Portal and from the U.S. Geological Survey (USGS). Therefore, all their terms of use and conditions apply here, too. These include the Terms of Use of According to the NYC Open Data terms of use, the originators of the individual datasets, i.e. the different departments of NYC administration, remain the owner of the data.

Note: RoofN3D is always in progress; the provided datasets will be improved, updated, and extended in the future.

Any scientific paper whose results are based on the RoofN3D training data must cite (Wichmann et al., 2018) and must contain the following acknowledgment:

The RoofN3D training data (Wichmann et al., 2018) was provided by the chair Methods of Geoinformation Science of Technische Universität Berlin and is available at

Related Projects

Building Roof Classification and Local Shape Estimation for Point Cloud Data Using Convolutional Neural Networks

We use different deep neural network architectures to test the potential of the RoofN3D dataset for different tasks towards 3D reconstruction in a deep learning framework. As the generating distribution of the data is different from those of natural images, we develop neural networks architectures that best accommodate geometric data, for building classification, estimation of the roof planes parameters and semantic segmentation.