One of the main goals of Computer Vision is the development of algorithms to perceive and understand our three-dimensional world. This task is extremely complex: the environments in which we usually operate are composed of a plurality of objects with the most diverse characteristics. Furthermore, 3D vision algorithms must be robust to noise and outliers that are inevitable when dealing with visual data.

GEOPRIDE (Geometric primitive fitting and decomposition for 3D shapes representation) aims to investigate the extent to which geometry can function as a 'universal language' that allows the complexity of the 3D world to be expressed. The alphabet of this language consists of a relatively small number of geometric primitives (lines, planes, quadrics, super-quadrics) that allow the complexity of a 3D scene to be described in simpler representations.

Instead of learning geometric primitives from examples in a training set, GEOPRIDE will develop robust automatic fitting procedures that operate on 3D data such as meshes or point clouds. In this way, the decomposition can be guaranteed both robustness against outliers and good generalisation capability, since the geometric estimation is agnostic with respect to the semantic classes to which the data belong. Finally, the research investigates how these extremely compact, accurate and scalable representations can be integrated into machine learning pipelines to process and analyse 3D data. In particular, the advantages of using geometric primitives in central applications for geometry processing such as shape segmentation, object retrieval and object recognition will be investigated.