Trading off salience and uncertainty in sampling a visual scene

摘要

When faced with a complicated visual scene many animals including humans attend to important regions in a systematic serial manner. The ability to orient rapidly towards an important region in a scene allows an organism to accomplish activities, such as navigation, foraging and detecting possible prey/mates. Developing a computational model of visual attention has long been of interest as such models enable artificial systems to acquire information efficiently from complex and cluttered environments. Current computational models attend to an important region (usually one which is maximally different from its immediate neighbours) and then inhibits future viewing of that region in order to facilitate distribution of visual attention. In this work we introduce the idea of an `uncertainty map', which works in conjunction with the existing idea of `saliency map' to drive the system's attention. We demonstrate the distribution of visual attention by our model in simulation. We show that despite its simplicity, our system distributes visual attention in a context-dependent manner which can be easily tuned to different environments.