Objectif
"Reliable recognition of thousands of object and action categories is today's key challenge in computer vision. Most contemporary approaches are based on supervised learning algorithms to train object classifiers. While manual annotation has become easier in recent years, it is still not scalable to a large set of categories. Moreover, as it is usually based on human language it does not reflect the visual characteristics of objects, but tries to establish high-level links that should actually be learned after appropriate visual features have been captured.
In this proposal, we aim at reducing the manual labeling effort by making use of the natural organization of visual data as it is provided by a video stream. In the same setting, we also aim at learning a more sophisticated structural representation of objects. Rather than manually specifying parts and attributes of objects that have a counterpart in language, we will seek correlated visual patterns by letting the data speak. Exploiting the natural arrangement of images in video and the inherent 3D scene structure is decisive, since weakly correlated images as obtained from photo collections might not contain rich enough relationship information.
We will also consider the active observer setting, i.e. where the camera can be moving. This allows extracting far more information, but also requires detailed control of the low-level and mid-level computer vision techniques involved, particularly motion estimation and tracking. The importance of these components is often underestimated in contemporary visual learning approaches.
Apart from the impact on the field of computer vision itself, the improved performance in visual recognition that we anticipate in this project has direct consequences for many important applications, particularly automotive systems and robotics, where the use of visual sensory input is more and more considered one of the most important components of future systems."
Champ scientifique
- humanitieslanguages and literaturegeneral language studies
- natural sciencescomputer and information sciencesartificial intelligencemachine learningsupervised learning
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringsensorsoptical sensors
- natural sciencescomputer and information sciencesartificial intelligencecomputer vision
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringrobotics
Appel à propositions
ERC-2011-StG_20101014
Voir d’autres projets de cet appel
Régime de financement
ERC-SG - ERC Starting GrantInstitution d’accueil
79098 Freiburg
Allemagne