Knowledge rules in support of point cloud semantic segmentation 

3 November 2022 | Online | 15:00 | Alessandro Daniele (FBK-DKM) and Eleonora Grilli (FBK-3DOM)


Deep Learning approaches have sparked much interest in the AI community during the last decade, becoming state-of-the-art in domains like pattern recognition, computer vision, and data analysis. In the geospatial and remote sensing field, the availability/unavailability of annotated training data is often a big obstacle, and most of the time, we have to deal with small samples and unbalanced classes. To overcome these problems, we introduce KENN (Knowledge Enhanced Neural Networks) within the 3D semantic segmentation pipeline. KENN is a library that allows neural network models to be modified by providing logical knowledge in the form of a set of logic constraints.

To give a practical example, in a point cloud semantic segmentation challenge, the rule which states that "poles"cannot be over a "building" is employed by KENN to avoid misclassification problems with antennas.

The work presented is the result of a collaboration between the 3D Optical Metrology Unit and the Data Knowledge and Management (DKM) Unit. 

Alessandro Daniele received his master's degree in Computer Science at Università degli Studi di Padova and his Ph.D. at Università degli Studi di Firenze. During his Ph.D. he worked in the Data Knowledge and Management (DKM) group at FBK, where he is currently a researcher in Neural-Symbolic Integration. His research topic is about the integration of logical knowledge with neural networks models, with a particular focus on relational domains.

Eleonora Grilli is a full-time postdoc researcher at the 3D Optical Metrology research group (3DOM) at the Bruno Kessler Foundation (FBK) in Trento. After her master's degree in Civil Engineering at the University of Bologna, she started and completed her Ph.D. at 3DOM under the supervision of Dr. Fabio Remondino, mainly focused on machine learning semantic segmentation applied to architectural and cultural heritage 3D datasets. Her current role involves the design of geometry and machine learning-based solutions in a wide range of fields, such as urban digitisation, digital agriculture and forestry, corridor mapping, and restoration projects.