D3.2 AI-based Indoor Scene Understanding v2
Vladimiros Sterzentsenko, Georgios Albanis, Nikolaos Zioulis, Vasileios Gkitsas, Antonis Karakottas, Petros Drakoulis, Pachalina Medentzidou, Alexandros Doumanoglou, Dimitrios Zarpalas, Werner Bailer, Stefanie Onsori-Wechtitsch, Hermann Fürntratt · Zenodo (CERN European Organization for Nuclear Research) · 2022
This deliverable documents the approaches for designing and developing the data-driven models that drive the ATLANTIS AI services backend. Their basic principle of operation is monocular inference from spherical panoramas which means that depending on the task at hand, these models will need to provide solutions for challenging, ill-posed problems. Yet the recent developments in data-driven models and the expanded availability of data are the drivers of such emerging services, upon which innovative tools and applications like the ATLANTIS authoring tool can be developed. In this updated version, the improved results from WP3 are presented which include works for scene layout and depth estimation, semantic segmentation, and a novel generative approach for Diminished Reality applied at spherical panoramas. These services have been integrated into the ATLANTIS application and assessed in the validation. Note that the deployment-related aspects (e.g., APIs, registration of the AR scene with the real room) are described in D4.2.