Seminar Series #1: Machine Learning with Structured (Geo)Spatial Information in Processing Unstructured 3D Point Clouds

Large-scale 3D point cloud data sets are increasingly available for next-generation buildings and city management, thanks to the exciting LiDAR and photogrammetry technologies in remote sensing. However, no “silver bullets” exist to resolve unstructured 3D point clouds representing these complex systems. Even the state-of-the-art 3D machine learning (ML) methods can hardly satisfy the classification or regression tasks of the point-level labels or object-level instances. In this talk, we will unveil the possibility of enriching generic ML with structured geospatial information, i.e., ML with (rather than of) geospatial information. First, we will present several general ways to chip structured domain knowledge into the paradigms of ML, including pre-processing, extra features, spatial attention, and post processing. Drawing from our recent work, we will showcase examples of detecting labels, global symmetries, local similarities, and objects in urban point clouds. Our dataset and experiments have demonstrated the unique and vital role of structured geospatial information in applying ML to built environment data sets, paving the way for long-term, innovative, smarter city applications.

Poster