A Model of Saliency-Based Visual Attention for Rapid Scene Analysis
IEEE Transactions on Pattern Analysis and Machine Intelligence
Decoding What People See from Where They Look: Predicting Visual Stimuli from Scanpaths
Attention in Cognitive Systems
Hi-index | 0.00 |
This paper presents an examination of the possible competition and cooperation that may take place in human visual attention, between the bottom-up saliencies incurred by photometric signatures and the top-down saliencies incurred by the primary context of a scene. It is found that the strength of the primary context of a scene represents a dominant guiding factor for determining the visual fixations for attention: in the case where there exists a strong context in a scene, the objects and/or regions that are tightly coupled with the context dominate for defining the saliencies that guide fixations for attention. It appears that, in human visual perception, a higher priority is assigned to the efficient understanding of a visual context than the direct response to photometric saliencies not supported by the context. The claims described above are derived from the experimental verification of the following conjectures: 1) There is a tendency for the bottom-up saliencies to be more significant when the context of the scenes observed is either weak or nonexistent. 2) For the scene of a strong context, the top-down context saliencies such as the objects and regions that are associated with understanding the present context tend to dominate over the bottom-up saliencies. 3) When the scene of a strong context includes both positive and negative saliencies, where the positive/negative contextual saliencies are referred to here as those saliencies significant for understanding the context yet well-expected/unexpected for the given context in terms of the prior knowledge, the negative saliencies are assigned a higher priority than the positive saliencies for attention.