Multi-Task Credible Pseudo-Label Learning for Semi-Supervised Crowd Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Self-training is a semi-supervised learning method for crowd counting, but pseudo-label noise in density maps hinders performance.
• Existing methods often ignore the multi-task relationships between density map regression and auxiliary tasks like segmentation.

Purpose of the Study:

• To develop a multi-task credible pseudo-label learning (MTCP) framework to address noise and improve semi-supervised crowd counting.
• To enhance feature representation learning by integrating density regression, binary segmentation, and confidence prediction.

Main Methods:

• A multi-task learning framework with shared feature extraction for density regression, binary segmentation, and confidence prediction.
• Data augmentation by trimming low-confidence regions in labeled data to reduce epistemic uncertainty.
• Direct generation of credible pseudo-labels for density maps in unlabeled data to decrease aleatoric uncertainty.

Main Results:

• The proposed MTCP framework significantly improves performance in semi-supervised crowd counting.
• The model demonstrates superiority over competing methods across four benchmark datasets.
• Credible pseudo-label generation for density maps effectively reduces noise and uncertainty.

Conclusions:

• The MTCP framework offers a robust solution for semi-supervised crowd counting by effectively managing pseudo-label noise and utilizing multi-task learning.
• The integration of auxiliary tasks and novel pseudo-labeling strategies enhances model accuracy and generalization.