The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLVI-4/W4-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVI-4/W4-2021, 91–96, 2021
https://doi.org/10.5194/isprs-archives-XLVI-4-W4-2021-91-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVI-4/W4-2021, 91–96, 2021
https://doi.org/10.5194/isprs-archives-XLVI-4-W4-2021-91-2021

  07 Oct 2021

07 Oct 2021

ALS POINT CLOUD CLASSIFICATION USING POINTNET++ AND KPCONV WITH PRIOR KNOWLEDGE

M. Kada and D. Kuramin M. Kada and D. Kuramin
  • Technische Universität Berlin, Institute of Geodesy and Geoinformation Science, Berlin, Germany

Keywords: Point clouds, 3D, Classification, Semantic Segmentation, City Modelling, Deep Learning

Abstract. In the practical and professional work of classifying airborne laser scanning (ALS) point clouds, there are nowadays numerous methods and software applications available that are able to separate the points into a few basic categories and do so with a known and consistent quality. Further refinement of the classes then requires either manual or semi-automatic work, or the use of supervised machine learning algorithms. In using supervised machine learning, e.g. Deep Learning neural networks, however, there is a significant chance that they will not maintain the approved quality of an existing classification. In this study, we therefore evaluate the application of two neural networks, PointNet++ and KPConv, and propose to integrate prior knowledge from a pre-existing classification in the form of height above ground and an encoding of the already available labels as additional per-point input features. Our experiments show that such an approach can improve the quality of the 3D classification results by 6% to 10% in mean intersection over union (mIoU) depending on the respective network, but it also cannot completely avoid the aforementioned problems.