Learning Pixel-Perfect 3D Vision and Generative Modeling

Informations projet

PIPE

N° de convention de subvention: 866435

DOI

10.3030/866435

Projet clôturé

Date de signature de la CE 7 Février 2020

Date de début 1 Septembre 2020

Date de fin 31 Août 2025

Financé au titre de

• EXCELLENT SCIENCE - European Research Council (ERC)

Coût total

€ 1 858 013,00

Contribution de l’UE

€ 1 858 013,00

Coordonné par

AALTO KORKEAKOULUSAATIO SR
Finland

Periodic Reporting for period 4 - PIPE (Learning Pixel-Perfect 3D Vision and Generative Modeling)

Période du rapport: 2025-03-01 au 2025-08-31

Résumé du contexte et des objectifs généraux du projet

PIPE seeks to advance computers' capability to reason about the physical world by looking at nothing but pictures. Moreover, PIPE seeks to enable this reasoning to stand on solid, physically meaningful quantities. Concretely, PIPE studies both the problem of understanding what shapes, materials, lights, objects, etc., appear in pictures of real-world scenes, as well as the problem of generative modeling, i.e. learning "a recipe" for the world.

Solutions to these fundamental problems in computer vision, computer graphics, and machine learning are relevant for a large number of applications from computer graphics, augmented and virtual reality to robotics; clearly, truly understanding the messy complexity of the real world must be necessary, e.g. for reliable, safe autonomous machines to emerge.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

PIPE seeks to advance the state of the art on three axes: (1) understanding the shapes in the world; (2) understanding the visual complexity of the surface materials in the world; and (3) learning generative computational recipes for things in the world. The most significant achievement since the beginning of the project is a result in the final direction that combines intuitive, controllable computational components with modern machine learning techniques. This enables photorealistic generative modeling in terms of physically meaningful concepts and quantities while requiring no human supervision.

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

PIPE is expected to provide breakthrough results in how well computer vision and generative algorithms can understand our three-dimensional world.

Overview of the approach in "A hybrid generator architecture for controllable face synthesis"