Weakly Supervised Visual Saliency Prediction.

Area of Science:

• Computer Vision
• Cognitive Science
• Artificial Intelligence

Background:

• Deep saliency models require extensive human fixation data, limiting their scalability and biological plausibility.
• Current supervised methods are annotation-intensive and do not fully capture visual attention mechanisms.

Purpose of the Study:

• To develop a weakly supervised deep saliency model grounded in cognitive theories of visual attention.
• To reduce the annotation burden for visual saliency prediction tasks.
• To enhance the understanding of underlying visual attention mechanisms.

Main Methods:

• Incorporated cognitive science insights as differentiable submodules into an end-to-end trainable framework.
• Integrated spatial and object-level semantics embedding, object relations, and a winner-take-all competition mechanism.
• Utilized a conditional center prior and novel loss functions based on image semantics, saliency priors, and self-information compression.

Main Results:

• The weakly supervised model achieved promising results in visual saliency prediction.
• Outperformed several fully supervised deep saliency models in experimental evaluations.
• Demonstrated the efficacy of cognitive-theory-driven components in saliency modeling.

Conclusions:

• The proposed weakly supervised approach significantly reduces annotation requirements for saliency prediction.
• Offers a more comprehensive and biologically plausible understanding of visual attention.
• Represents a key advancement in developing efficient and effective visual saliency models.