Program

Session I: Categorization

Coffee Break (10:40-11:10)

Session II: Attention

Motivation

Computer vision is gaining importance in the fields of artificial cognitive systems and robotics, due to the progress achieved in the last years in object localization, categorization and scene analysis as well as its low cost and versatility. From robot localization to manipulation, the integration of state of the art vision algorithms into robotic systems is a success story. Still, research in the two fields are largely separated. Vision has been traditionally studied using a reductionistic approach. This trend is even more pronounced now, as computer vision moves towards internet vision, i.e. computer vision methods for browse, classify and understand images found on the Web. We argue that issues such as multi-cue integration, embodied categorization and situated attention should be studied in the context of systems. The goals of this workshop are to document the progress of the relatively young field of cognitive computer vision and systems, to bring together the researchers working and interested in this field and giving them a platform to discuss the most recent advances and what are the research challenges that is timely to attack today.

Topics

The meeting will particularly focus on two main issues: categorization and attention, and how they are declined for a situated agent, or for a web-agent engaged with millions of images available on the Web.

• Categorization: The capability to categorize objects on the basis of their visual appearance is one of the crucial cognitive abilities that enable humans to understand the outside world and interact with it. Providing an autonomous robot with the same capability is a major scientific challenge. The computer vision community has achieved impressive results in this field recently, but these results are not easily exploited by the robotics community. The current mainstream approach in computer vision performs categorization from collections of static images, typically acquired on the Web. A very recent trend focuses on the ability to categorize 10,000 object categories, representing objects on the basis of only one type of visual feature. New methods are needed to enable abstractions and effective categorization, keeping into account the 3D structure of object categories, their associated affordances and how embodiment, context and task affects modeling and learning for an autonomous agent.
• Attention: 30 or so years ago attention was compelled by the hard constraints of the limited computational resources computer vision had to deal with. But nowadays, in the era of the large database and machine learning and brute force solutions, powerful statistical classifiers may yield impressive single task performance. The drawback is that they do not lead to representations suitable for reasoning and acting in a real world as it is required by an autonomous cognitive agent. Further, they ignore the fact that an autonomous agent world is dynamic, unpredictable, and not describable by ten thousands of objects. Even the databases on which these systems are tested require re-consideration so as to ensure they are statistically valid for the tasks at hands, and are not biased. Still, results are impressive and only promise to become more so with computer power increases and declining costs. Still, the goal for vision systems is really the same as it was 30 years ago. We want systems to be robust to the whole variability in the visual world, to the way we view the world and to the knowledge we have of the world, we want them to be flexible and not be single-task systems, we want them to do more than image classification, and visual reasoning and problem solving remain important unsolved problems. We want them to properly deal with the unexpected or a not-previously-viewed scene. Attention is that capacity that helps optimize the search processes inherent in perception, cognition and action, thus reducing the computational load of an agent. The spectrum of attentive behavior is broad and goes far beyond the simple region-of-interest functions most common in today's systems. When embodied in an agent, attention controls active sensing and action and promises to enable the kind of flexible and robust systems that have been the goal of computer vision since its earliest days.