Learning a Combined Model of Visual Saliency for Fixation Prediction

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Numerous saliency models exist, each excelling on specific image types but lacking generalizability.
• Individual models struggle with arbitrary images and large datasets due to single-hypothesis limitations.

Purpose of the Study:

• To improve general saliency detection accuracy by fusing multiple state-of-the-art models.
• To investigate the effectiveness of score-level fusion using para-boosting learning strategies.

Main Methods:

• Saliency maps from various models were treated as confidence scores.
• These scores were fed into para-boosting learners (SVM, AdaBoost, KDE) for final map generation.
• Compared para-boosting with traditional fusion methods (Sum, Min, Max).

Main Results:

• Score-level fusion significantly outperformed individual saliency detection models.
• The proposed fusion method narrowed the performance gap with human observers.
• Consideration of fewer models reduced computational cost.

Conclusions:

• Para-boosting score-level fusion is an effective strategy for enhancing general saliency detection.
• This approach offers a robust solution for improving performance on complex image datasets.
• Fusion mitigates the limitations of single-hypothesis saliency models.