Programme

"Dynamic optical 3D metrology for wind energy applications"

Thomas Luhmann - Institute for Applied Photogrammetry and Geoinformatics, Jade University of Applied Sciences (Germany).

The keynote will present recent developments and projects in the field of high-dynamic photogrammetry and scanning techniques for the acquisition of wind energy systems in wind tunnel and in situ experiments.

"SfM versus marker-based close-range photogrammetry: distinctions and potential for an integrated approach"

Clive Fraser - Professorial Fellow, Dept. of Infrastructure Engineering, University of Melbourne (Australia)

Within industrial and engineering photogrammetry, the utilisation of targets/markers is near universal, whereas the adoption of automated, targetless orientation and object reconstruction is becoming increasingly more widespread across other application domains. This naturally poses the question of why so-called Structure-from-Motion (SfM) techniques are yet to be employed to any significant extent in large-scale vision metrology, in spite of the obviously appealing prospect of dispensing with artificial signalisation, be it through physical targets or projection schemes. This presentation reviews the principal distinctions between the SfM and marker-based approach, with the emphasis being upon measurement accuracy, reliability and practicability. Also discussed will be the potential for an integrated approach employing both SfM and targets, with a project example of where such an integrated network orientation and point determination approach proved to be optimal in terms of both accuracy and productivity. The project involved deformation measurements of an inflatable antenna on the International Space Station.

"How Does Context Help for the Semantic Interpretation of 3D Point Clouds?"

Martin Weinmann

Institute of Photogrammetry and Remote Sensing, Karlsruhe Institute of Technology - KIT, (Germany)

3D data in the form of irregularly distributed 3D points serve as the basis for a diversity of applications such as geographic information systems, building/city modeling, construction site monitoring, rail infrastructure mapping, urban planning, autonomous driving, or creation of digital twins. Many of these applications rely on the automatic interpretation of such 3D data to reason about properties of the scene and/or objects therein. The reasoning, in turn, can be tackled by using traditional approaches that rely on handcrafted features, but also by using modern deep learning approaches that rely on the learning of suitable features to solve a particular task. While deep learning has been proven beneficial for a diversity of tasks addressing image interpretation, the transfer of deep learning to 3D point clouds is not straightforward due to the unstructured nature of the underlying data. This presentation will exemplarily focus on the semantic segmentation of 3D point clouds acquired from terrestrial and airborne platforms via laser scanning, and provide a survey addressing the three main strategies to achieve this: (1) pointwise classification, (2) contextual classification, and (3) deep-learning-based classification. In this scope, the strengths and weaknesses of these strategies will be discussed, and the common foundations and conceptual differences will be highlighted.