Application of Ontologies for Semantic Scene Segmentation and Object Recognition

Abstract:
Semantic scene segmentation is a key block of many human-robot collaboration applications. Existing methods yield good results in lab environment both on a class and instance level. Additionally, some work was done in the development of tools to simplify the task of teaching the semantics. However, this resides a tedious task that doesn’t adapt well, and lacks reusability. We propose a novel pipeline design that allows to build shareable, sensor agnostic models for objects that can later be used to identify scene segments. The result is an ontology IRI that can be shared and to which much more information can be linked. The method is based on traditional 3D vision techniques and deep learning classifiers, combined together using a RDF topology of OWL concepts.