important dates

April 15, 2014
Paper submission

April 30, 2014
Notification of acceptance

June 1, 2014
Workshop at ICRA 2014

Abstract

Visual search in complex scenes remains a challenging problem as robots start to move into less structured factory floors and homes. The recent move to RGB-D processing has helped simplify some vision problems, but also led to increased amounts of data to be processed. In these scenarios robots will always face situations, no matter how optimised and parallelised the vision algorithms, where the visual input is too overwhelming to process within the time constraints of the task at hand.

Visual attention is an integral part of human vision that allows humans to process the relevant parts of a scene from a large sensory stream within the harsh time constraints of real world tasks. Robotic vision research has realised the importance of attention mechanisms, and many systems use some form of attention to direct the processing to the parts of the sensory input that are potentially of interest, such as objects emerging from a supporting plane. The last decade has also seen an increasing number of principled approaches on making attention mechanisms an integral part also of robotic vision systems.

This poses a number of interesting challenges, such as selection of suitable mechanisms from a large body of work on biologically plausible attention models, fusing general purpose bottom-up and task-driven top-down attention mechanisms, how to use attention in a given task setting to guarantee robot performance within specific time constraints, how to integrate attention algorithmically into the various processes of an integrated robotic architecture, and coping with conflicting goals such as paying attention to avoid collision with humans while looking at task relevant target locations.

In order to tackle these problems, the robotics community needs continuous collaboration and communication with researchers on the psychology and neurology of biological vision systems. While the algorithms and architectural implementations will necessarily differ, insights from biological vision are highly relevant in this still poorly understood field of robot vision.