Animate Vision in a Rich Environment

Most research in computer vision has been directed towards minimalistic approaches, in which problems are addressed on how properties of the environment can be computed from as little information as possible. Although such approaches may be scientifically well motivated they have only resulted in limited progress towards our understanding of seeing systems. Ballard, Bajcsy and others have pointed out the importance of vision being an active process which is tightly connected to behaviors. We support this thought and also propose that utilizing that the world is rich on information is essential. We develop this idea to show how attention and figure-ground segmentation by an active observer using multiple cues can be separated from analyzing and recognizing what is seen in a consistent way. Continuous operation over time and early use of three dimensional cues are important in this context. We illustrate our proposed approach by some experiments on a real-time active system. 1 Introduction Vision is a sense by which seeing creatures acquire information about a dynamically changing environment and thereby guide many of their behaviors and actions. Computer vision research aims at understanding and developing computer based systems with such capabilities. Despite extensive efforts for more than three decades we still seem to be very far from such a goal. Although there exists ample knowledge of how information about the environment can be computed from visual cues, we see little progress towards what can be called seeing systems. has pointed out that the major reason for this is that vision, as we know it from biology, is an active process and that traditional computer vision approaches take no heed of this fact. Ballard [1989, 1991] analyses this discrepancy and its consequences. Here we will further Ballard's arguments and discuss some additional issues which we believe are crucial. We will also report on recent progress towards the realization of animate vision systems. Emphasizing the strong ties between vision and behaviors Ballard particularly considers the need for gaze control and what he terms "quickly computable features"1. The point that we want to stress in this context is that the real world is rich on information and that a multitude of such features can be computed when needed. We will argue that this suggests a paradigm of attentional mechanisms coupled to possibly independent mechanisms for deriving scene characteristics and information about objects. Notably this implies that the environment itself influences what should …