Weakly Supervised Human Fixations Prediction.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Cognitive Science

Background:

• Predicting human eye fixations is crucial for multimedia applications but current models struggle with object-level scene understanding and explicit spatial relationships.
• Existing saliency models often depend on object detectors, limiting them to predefined categories and failing to capture nuanced spatial interactions.

Purpose of the Study:

• To develop a weakly supervised fixation prediction model that leverages image labels for enhanced accuracy.
• To hierarchically discover objects and model their spatial configurations for more robust eye movement prediction.

Main Methods:

• The proposed method samples superpixels to generate object-level graphlets (oGLs) via random walking.
• A manifold embedding algorithm encodes image labels into oGLs, computing object response maps.
• Spatial-level graphlets (sGLs) are introduced to model inter-object positions, integrating eye-tracking data for fixation prediction.

Main Results:

• The model successfully discovers objects and their spatial configurations hierarchically.
• Weakly supervised learning with image labels significantly improves fixation prediction accuracy.
• Experimental results demonstrate the proposed method's superiority over state-of-the-art approaches.

Conclusions:

• The proposed weakly supervised fixation prediction method effectively addresses limitations of conventional approaches.
• Hierarchical discovery of objects and spatial relationships enhances the prediction of human eye movements.
• This approach offers a more comprehensive and accurate solution for eye fixation prediction in multimedia contexts.