Is there a shape to the attention spotlight? Computing saliency over proto-objects predicts fixations during scene...

Area of Science:

• Cognitive Psychology
• Computational Neuroscience
• Computer Vision

Background:

• Visual attention guides information processing, with its spatial distribution often modeled as Gaussian.
• Previous models did not fully account for the potential shape of attentional focus.

Purpose of the Study:

• To propose and test a novel model of visual attention based on proto-objects (POs).
• To investigate whether attention's spatial routing is better described by shape than by a Gaussian distribution.
• To compare the predictive power of a PO-based model against traditional saliency and object-based models.

Main Methods:

• Developed an image-computable model that averages saliency over proto-objects (POs).
• Collected human free-viewing data (fixation density maps and scanpaths) for 384 natural scenes from 12 participants.
• Compared model predictions against empirical fixation data and against a model using hand-segmented object labels.

Main Results:

• The PO-based attention model significantly outperformed saliency models lacking shape considerations in predicting fixation density and scanpaths.
• The PO model's predictive accuracy for fixations was comparable to a model utilizing explicit object segmentation.
• Proto-objects (POs) provide a quantifiable representation of the shape of attention.

Conclusions:

• The spotlight of attention is not necessarily Gaussian but possesses a definable shape.
• Proto-objects (POs) offer a robust framework for understanding and modeling the spatial dynamics of visual attention.
• This research advances computational models of attention by incorporating object-level spatial properties.