Seminar: Neural Methods for 3D World Understanding

Understanding 3D structure is fundamental to scientific discovery, yet current neural methods struggle to capture the geometric and topological relationships that define complex 3D environments. This talk presents work on developing neural architectures that can effectively learn 3D world structure across multiple scales and applications.

The talk introduces two complementary research directions. First, large-scale analysis methods demonstrate how graph neural networks and attention mechanisms can extract semantic understanding from massive 3D datasets. BuildingNet, a semantically annotated buildings dataset, illustrates how neural architectures can learn architectural structure and relationships. Work on remote sensing applications shows how autoregressive sequence modeling can extract building footprints from satellite imagery, casting spatial understanding as a language modeling problem.

Second, the talk discusses recent advances in generative approaches that employ structural understanding to create controllable visual content. Skeleton-based generative models and 3D-conditioned image synthesis methods demonstrate how explicit structural representations enable more precise control over generated content.

Throughout this work, the research also addresses the computational challenges of processing massive 3D datasets using high-performance computing infrastructure, including the management of the Prometheus AI supercomputer. Looking forward, the talk explores potential applications of structure-aware AI in scientific computing, examining how explicit geometric representations could complement existing computational approaches across various domains.

Dr. Melinos Averkiou is an Associate Research Professor and Group Leader of the Visual Computing Group at CYENS Centre of Excellence in Cyprus. He received his PhD in Computer Science from University College London in 2015, and has held research positions at Stanford University, SIAT China, and the University of Cyprus.

His research focuses on developing neural methods for understanding and generating visual content, with applications spanning computer vision, urban planning and remote sensing. He also manages Prometheus, CYENS' peta-scale AI supercomputer.