on Object-based Visual Attention for Robots
Yuanlong Yu, George K. I. Mann, Raymond G. Gosine · 2009
The selectivity of visual attention mechanism is influenced by bottom-up competition and top-down biasing. This paper presents an object-based visual attention model which simulates top-down influences. Five components of top-down in- fluences are modeled: learning of object representations stored in long-term memory (LTM), deduction of task-relevant feature(s), estimation of top-down biases, mediation between bottom-up and top-down fashions, and object completion processing. This model has been applied into the robotic task of object detection. Experimental results in natural and cluttered scenes are shown to validate this model. I. INTRODUCTION Attention mechanism sheds some light on developing robotic visual perception since it can filter out irrelative information whereas limit processing to items that are relevant to the present task (3). Two assumptions have been proposed to model attention: space-based and object-based. Duncan's inte- grated competition (IC) hypothesis (4) shows that it produces a competitive advantage over the whole object by directing attention to an attribute of one object. The attribute could be represented by a spatial location, a feature dimension or a part. Besides, some computational factors also lead us to develop an object-based attention model for robots: 1) Attending to an object can provide more useful information (e.g., shape, size and spatial region) for robots to produce appropriate actions than to a spatial location; 2) Estimation of attentive activation at the object level by accumulating contributions of all components within an object is more robust to noise than at the spatial location level without accumulation; 3) Object- based attention mechanism is the only way to realize top-down biasing if the conspicuous feature is represented by global features (e.g, shape). Attentional selection is mainly modulated by two factors (2): bottom-up competition and top-down biasing. Bottom- up competition performs contrast in the spatial context to guide attention towards local inhomogeneous objects, while top-down biasing modulates attentional competition in favor of task-relevant objects. Both contributions are combined together to decide which object is attended to.