Rethinking Self-Training for Semi-Supervised Landmark Detection: A Selection-Free Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Self-training is a semi-supervised learning technique where pseudo-labels are crucial for managing confirmation bias.
• Existing self-training methods struggle with landmark detection due to data bias in pseudo-labels, difficulty in threshold selection, and lack of confidence scores from coordinate regression.

Purpose of the Study:

• To propose a novel self-training method for landmark detection that addresses the limitations of existing approaches.
• To develop a method that does not rely on explicit pseudo-label selection, mitigating issues like data bias and noisy labels.

Main Methods:

• Introduced Self-Training for Landmark Detection (STLD), a method employing a task curriculum to manage confirmation bias.
• Utilized pseudo pretraining for better model initialization and shrink regression for coarse-to-fine pseudo-label leveraging in later stages.

Main Results:

• STLD demonstrated consistent outperformance over existing methods across facial and medical landmark detection benchmarks.
• The proposed method proved effective in both semi-supervised and omni-supervised learning scenarios.

Conclusions:

• STLD offers a robust solution for landmark detection by effectively handling confirmation bias without explicit pseudo-label selection.
• The task curriculum approach, combined with pseudo pretraining and shrink regression, significantly enhances model performance in semi-supervised landmark detection.