SIMULACRON: From Camera Observations to Physical Simulations of the 3D World

Descripción del proyecto

Hacia un novedoso concepto de interpretación dinámica de imágenes en la visión artificial

La visión artificial se encuentra en el punto de intersección entre la informática, las matemáticas, la ingeniería y la física. Se centra en imitar parcialmente la complejidad del sistema de visión humano y es uno de los tipos más potentes de inteligencia artificial. El proyecto SIMULACRON, financiado con fondos europeos, desarrollará un concepto más profundo de interpretación dinámica de imágenes en la visión artificial. Se centrará en deducir la física subyacente (masas, elasticidad, cantidad de movimiento y fuerzas) y en una simulación de la acción observada directamente a partir de los vídeos. En particular, el proyecto desarrollará algoritmos para una modelización de formas deformables y métodos variacionales para inferir simulaciones físicas mediante vídeos. Por último, SIMULACRON encabezará una transición de la inferencia de geometría tridimensional a la inferencia de simulaciones físicas.

Objetivo

Despite their amazing success, we believe that computer vision algorithms have only scratched the surface in terms of understanding our world from images. While most research on 3D reconstruction has been concerned with recovering the surface geometry and reflectance, SIMULACRON is focused on inferring the underlying physics (masses, elasticity, momenta, forces, etc.) and a simulation of the observed action directly from videos.

This not only provides a more profound understanding of the observed phenomena, but it also allows us to interpolate and extrapolate complex actions far beyond the observation: The inferred physical simulation can be employed for space-time super-resolution and for predictions into the future.

SIMULACRON covers three lines of research:

A) We will develop algorithms for deformable shape modeling. We will explore suitable representations of 3D shape and its evolution that enable the efficient computation of shape deformation, correspondence, interpolation and extrapolation. These techniques will form the basis for inferring physical simulations in parts B and C.

B) We will develop variational methods for inferring physical simulations from videos. We will compute a reference shape and simulation parameters that generate the shape deformation that is most consistent with the observations.

C) We will develop learning-based approaches for inferring physical simulations from videos. We will pursue two alternative approaches: First, we will generate synthetic training data by simulating deformable shapes and the associated camera observations. Second, we will devise self-supervised techniques for learning from real data without requiring labeled training data.

By shifting from inference of 3D geometry to inference of physical simulations, SIMULACRON will give rise to a more profound notion of dynamic scene understanding in computer vision, robotics and beyond. We believe that we have the necessary competence to pursue this project.