Predicting Human Attention using Computational Attention

Zhibo YangSounak MondalSeoyoung AhnGregory ZelinskyMinh HoaiDimitris Samaras

Mar 2023

摘要

Most models of visual attention are aimed at predicting either top-down orbottom-up control, as studied using different visual search and free-viewingtasks. We propose Human Attention Transformer (HAT), a single model predictingboth forms of attention control. HAT is the new state-of-the-art (SOTA) inpredicting the scanpath of fixations made during target-present andtarget-absent search, and matches or exceeds SOTA in the prediction of tasklessfree-viewing fixation scanpaths. HAT achieves this new SOTA by using a noveltransformer-based architecture and a simplified foveated retina thatcollectively create a spatio-temporal awareness akin to the dynamic visualworking memory of humans. Unlike previous methods that rely on a coarse grid offixation cells and experience information loss due to fixation discretization,HAT features a dense-prediction architecture and outputs a dense heatmap foreach fixation, thus avoiding discretizing fixations. HAT sets a new standard incomputational attention, which emphasizes both effectiveness and generality.HAT's demonstrated scope and applicability will likely inspire the developmentof new attention models that can better predict human behavior in variousattention-demanding scenarios.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答

被引用

笔记

问答

领域