• 3d scene understanding
• computer vision

The ultimate goal of this thesis work was to structure a framework for performing 3D open vocabulary scene understanding in an innovative manner compared to the currently existing state-of the art solutions in this context. The pre-existing solutions, in fact, employ not only three-dimensional data related to the input environments, but also a wide range of RGB-D images of the same scenes from different angles, which enable more comprehensive and accurate semantic identification and understanding of segmented objects.
Differently, the framework we hypothesized aims to exclusively utilize environments in the form of three-dimensional point clouds, without additional usage of other data typologies to support task completion: to achieve this, a synergistic employment of two distinct components is necessary, which will constitute a unified data flow that starts from the real environment and concludes with its semantic comprehension.