Learning a saliency map using fixated locations in natural scenes.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Psychophysics

Background:

• Computational saliency models, inspired by primate vision, predict visual attention by integrating feature maps.
• Existing models often use fixed weights, potentially limiting their accuracy in predicting human eye movements on natural scenes.

Purpose of the Study:

• To develop an improved computational saliency model by learning optimal weights for feature maps using real eye-tracking data.
• To investigate the influence of different feature channels (face, orientation, color, intensity) on visual attention.
• To model and account for the central bias observed in human fixation patterns.

Main Methods:

• Employed a least squares technique to learn feature map weights from four natural scene eye-tracking datasets.
• Assessed model performance using metrics beyond the standard area under the ROC curve.
• Incorporated both time-varying and constant factors to model the central fixation bias.

Main Results:

• Learned weights varied across datasets, with face and orientation channels generally being more important than color and intensity.
• Inter-subject differences in learned weights were found to be negligible.
• The developed model demonstrated superior performance compared to several state-of-the-art saliency algorithms.

Conclusions:

• The data-driven approach significantly enhances saliency model performance by adapting feature map contributions.
• The model's simplicity facilitates application in psychophysics, physiology, and real-world computer vision systems.
• This work provides a more accurate and adaptable method for predicting visual attention.