A Qualitative Analysis Strategy Towards AI-Enabled 3D City Reconstruction
Abstract:
3D urban scene reconstruction is a difficult problem in computer graphics, mainly due to unavoidable noise in real scene datasets and scalability. Whilst the state-of-the-art is able to produce reconstructions of various qualities, very few reconstruction approaches incorporate noise and clutter treatment strategies. Funded by Ordnance Survey, in this dissertation, we identify key areas of potential innovation in order to achieve large-scale and denoised 3D models of cities. To determine these areas of innovation, qualitative data from stakeholder engagement activities have been used to guide a literature survey on recent techniques in 3D reconstruction. The KJ method, also known as Affinity Diagrams, has been employed to organise and visualise the qualitative data collected. The steps of the method were closely followed where appropriate, and the process of arriving at the finalised diagram is shown. The literature surveyed is then mapped on the Affinity Diagram to identify the strengths and gaps of the literature with respect to our stakeholder’s requirements and workflow practices. Following this approach, the main goal of our stakeholder has been defined as a multiclass reconstruction of urban scenes. Along with the 3D reconstructions themselves, two more areas for future investigation are identified, which combined promise to achieve our stakeholder’s goal. The first area identified is semantic capabilities, which are capabilities referring to the ability to understand different objects and shapes within the scene. The second area identified is human-centric quality assurance, where the hypothesis is that the incorporation of people within the pipeline to be developed will greatly aid in achieving scalable multi-class reconstructions of that scale.
Download Andrea's Thesis