Architecting an Artificial Intelligence Ontology System - Ann Clark

Artificial Intelligence (AI) requires large amounts of labelled training data for models to be accurately trained. In semantic segmentation, each pixel is assigned a value based on the class or classes it belongs to. On the other hand, an ontology is a collection of classes, their properties, and the relationships between them. Semantic segmentation, then, lends itself naturally to integration with an ontology, where insights from direct and indirect relationships between classes can be used to enhance training data.

Recently, Nearmap introduced an ontology system which integrates directly into our proprietary labelling software, linking labelled data directly to definitions. Preliminary work on the system has already helped identify conflicts between labels in the production dataset.

Along the way, a few challenges unique to managing Nearmap’s large geospatial dataset had to be overcome. This presentation will discuss the solution which handles reproducibility, version tracking, multiple perspectives in source imagery, and the storage of image examples as part of the class definitions.

Ann Clark: AI Ontologist, Nearmap company, Australia