Thursday, 28th November

Abstract: 3D Computer Vision has recently witnessed a surge of interest from the ML and CV research community, due to the progress that recently introduced concepts such as neural representations, foundational models, and diffusion models enabled for many traditional 3D Computer Vision tasks. In this talk, we will focus in particular on the capability of understanding and synthesising 3D scenes and objects, which is a key component of applications in the space of Augmented/Mixed Reality and Robotics. We will look at three tasks in 3D Computer Vision that are fundamental components for these applications while being highly influenced by the aforementioned concepts: novel view synthesis, 3D semantic segmentation and 3D asset generation. For each of these three tasks, we will first understand some important practical limitations of current approaches. We will then walk through some solutions I recently explored with my team and designed to overcome such limitations, which include robust novel view synthesis, open set 3D scene segmentation and realistic 3D asset generation.

https://federicotombari.github.io/