Exploiting Unlabeled Data in CNNs by Self-Supervised Learning to Rank.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Collecting labeled data for machine learning is costly and time-consuming.
• Self-supervised learning leverages abundant unlabeled data through auxiliary tasks.
• Ranking can serve as an effective proxy task for regression problems.

Purpose of the Study:

• To develop a self-supervised learning framework using ranking for regression tasks.
• To propose an efficient backpropagation technique for Siamese networks.
• To improve performance in Image Quality Assessment (IQA) and Crowd Counting.

Main Methods:

• Utilized ranking as a proxy task for regression.
• Developed an efficient backpropagation method for Siamese networks.
• Generated ranked image sets from unlabeled data for IQA and crowd counting.
• Integrated supervised regression with self-supervised ranking objectives.

Main Results:

• Achieved state-of-the-art results in both IQA and crowd counting.
• Demonstrated significant performance improvements by combining labeled and unlabeled data training.
• Showcased that network uncertainty on the proxy task effectively measures data informativeness.
• Validated the framework's ability to reduce labeling effort by up to 50% through active learning.

Conclusions:

• Self-supervised learning via ranking is a viable and effective approach for regression tasks.
• The proposed Siamese network optimization enhances training efficiency.
• The framework offers a powerful method for improving computer vision tasks with limited labeled data.
• Uncertainty-based active learning significantly reduces data annotation costs.