Despoina Paschalidou

Research Abstract:

Humans develop a common-sense understanding of the physical behavior of the world, within the first year of their life. For example, we are able to identify 3D objects in a scene, infer their geometric and physical properties, even when they are partially visible, predict physical events in dynamic environments, reason about intentions and decide our actions based on our interactions with our surroundings. While the above tasks are fairly easy for the human brain, existing computer vision algorithms struggle to form such high-level reasoning. This is anticipated since reliably perceiving the physical world involves jointly reasoning about static and dynamic elements in a scene i.e. their semantic composition, spatial arrangement etc., as well as interpreting them independently i.e. their physical properties, shape, appearance, functionality etc. The need to efficiently capture the compositional nature of the physical world, gave rise to techniques that seek to jointly reason about all elements in a scene. My work so far focused on developing compositional representations for 3D objects and scenes that can be used for a great variety of tasks such as 3D reconstruction of objects and scenes, generative models of objects and scenes as well as videos. In the future, I would like to develop expressive representations that are equipped with common sense understanding and can efficiently capture more complex concepts such as interactions between humans, interactions between humans and objects, as well as interactions between several objects. I strongly believe that exploring such representations both in the context of generative as well as perceptual tasks will greatly benefit several robotics and AR/VR applications. Moreover, I am also very passionate about developing generative pipelines that can synthesize controllable 3D worlds with multiple static and dynamic components. I think one important aspect of such models is to keep humans in the loop, namely allow people to explore their creativity through generative AI.

Bio:

Despoina Paschalidou is a Postdoctoral Researcher at Stanford University working with Prof. Leo Guibas at the Geometric Computation Group. Prior to this, she did her PhD at the Max Planck Institute for Intelligent Systems in Tubingen and the Computer Vision Lab in ETH Zurich, under the guidance of Prof. Andreas Geiger and Prof. Luc van Gool. She received her Diploma in Electrical and Computer Engineering from the Aristotle University of Thessaloniki. During her PhD, she spent one wonderful year working with Prof. Sanja Fidler at NVIDIA Research, where she was involved in the development of several tools for 3D Generative AI. Moreover, she also had the pleasure to intern at FAIR under the guidance of Prof. Andrea Vedaldi and Dr. David Novotny, where she worked towards developing 3D-aware representations conditioned on videos.