Feature Map Quality Score Estimation Through Regression.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Estimating visual quality of feature maps is crucial for active vision applications.
• Existing methods using compactness fail on blurry maps like fixation maps.
• A new regression-based approach is needed for accurate feature map quality assessment.

Purpose of the Study:

• To develop a robust method for estimating the visual quality of feature maps, specifically fixation maps.
• To address limitations of previous compactness-based methods, especially for blurry maps.
• To improve human fixation prediction by dynamically integrating feature maps based on quality.

Main Methods:

• Feature map quality estimation treated as a regression problem.
• Extraction of local, global, geometric, and positional features.
• Utilization of a random forest regressor model trained on diverse feature maps.

Main Results:

• The proposed model accurately estimates the quality of bottom-up, target, and contextual feature maps.
• The approach demonstrates effectiveness on a large dataset of challenging outdoor images.
• Dynamically selecting feature maps based on estimated quality enhances fixation prediction accuracy.

Conclusions:

• The developed regression model provides a reliable method for assessing feature map visual quality.
• This quality estimation improves the integration of multiple feature maps for better human fixation prediction.
• The approach offers a significant advancement in active vision and computational attention models.