The role of visual and semantic similarity in the allocation of visual attention

Abstract

Contemporary models of attentional deployment can be categorised as representing either a more top down/endogenous or bottom-up/exogenous approach. Recently, within the scope of visual attention, the role of semantic information in attentional guidance has garnered some significance (Wu et al., 2014; Henderson and Hayes, 2017) as opposed to traditional saliency based modelling. This project conducted three research tasks investigating the roles of visual similarity and semantic similarity in the allocation of exogenous and endogenous attention. Visual similarity was defined by the SSMIM Matlab function which calculates a score of similarity in terms of three low level features, i.e. contrast, luminance and structure. Semantic similarity was defined by the word2Vec framework (Mikolov et al., 2013) which quantifies context vectors. While the first two tasks on exogenous attention did not yield significant results, the third task on endogenous attention found significant effects of visual and semantic similarity in the allocation of attention, hinting at possible underlying mechanism in visual processing.