Learning Saliency From Single Noisy Labelling: A Robust Model Fitting Perspective.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep neural networks advance visual saliency prediction but require extensive pixel-wise annotated data.
• Pixel-wise annotation is labor-intensive, posing a significant bottleneck for large-scale dataset creation.

Purpose of the Study:

• To develop a unified framework for learning saliency prediction from single noisy labels.
• To address the challenge of identifying clean labels within a noisy dataset for robust model training.

Main Methods:

• Formulated deep saliency prediction from noisy labels as robust network learning, leveraging robust model fitting theory.
• Exploited model consistency across iterations to distinguish inliers (clean labels) from outliers (noisy labels).

Main Results:

• The proposed framework achieves comparable saliency prediction performance to state-of-the-art fully supervised methods.
• Demonstrated tangible improvements even when treating ground truth annotations as noisy labels, outperforming existing methods.

Conclusions:

• The developed framework offers an effective solution for learning visual saliency from easily obtainable noisy labels.
• This method reduces the reliance on laborious pixel-wise annotations, making saliency prediction more accessible and efficient.