EPSRC Doctoral Training Partnership PhD: 3D Shape Analysis using Deep Learning

3D shapes have proliferated in recent years, due to the availability of low-cost 3D cameras such as Kinect and Occipital Structure sensors, and their wide applicability in manufacturing design, entertainment industry, digital earth, 3D printing, and virtual reality etc. Online databases such as Trimble 3D Warehouse contain millions of 3D shapes. It is an increasing challenge to describe complicated 3D shapes effectively and analyse shapes at a high-level, e.g. to recognise objects in a scene or to find reusable parts for manufacturing design.
Existing shape analysis techniques are largely derived from geometry such as differential geometry which is mathematically elegant but may not capture the high-level perception or semantics of 3D shapes. The huge amount of 3D shapes available in online repositories opens up a new route to shape analysis using data-driven machine learning based methods. Both supervisors have extensive experience of data-driven geometric modelling, with recent research covering data-driven shape editing, shape interpolation, scene modelling, bas-relief generation etc. Such state-of-the-art results provide a solid foundation for further research in this direction.
Deep learning has shown to be an effective approach in image analysis, achieving state-of-the-art performance in many tasks. Comparatively little research has considered deep learning for 3D shape analysis, and majority of them treat 3D shapes as volumetric grids or 2D projections, which either incur a very high computational cost, or fail to capture the full characteristics of 3D shapes, thus limiting their applicability. There is great potential to extend image based deep learning techniques to general surface-based shape analysis, where the underlying representation is not regular grids but general triangular mesh surfaces. Similar to the image counterpart, this can lead to more effective shape descriptors than the hand-crafted ones that are currently widely used, as well as leading to more effective techniques for high-level shape recognition.