Crowd Counting Using Meta-Test-Time Adaptation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning is vital for efficient crowd counting.
• Current test-time adaptation methods often require extensive training and unannotated data.
• Unsupervised domain adaptation is the dominant approach, demanding significant resources.

Purpose of the Study:

• To introduce CrowdTTA, a novel meta-test-time adaptive crowd counting approach.
• To enable crowd counting models to adapt to unknown test distributions more effectively.
• To reduce the reliance on large amounts of unannotated data for new target domains.

Main Methods:

• Integrates test-time adaptation with meta-learning for crowd counting.
• Introduces uncertainty via dropout layers to generate pixel-level pseudo labels.
• Employs a dual-level optimization process: inner self-supervised update and outer ground-truth update.

Main Results:

• Demonstrates general adaptive capability across diverse datasets with varying crowd densities and scales.
• Outperforms various supervised learning and domain adaptation methods.
• Effectively adapts models to unknown test distributions with improved performance.

Conclusions:

• CrowdTTA offers an efficient and adaptable solution for crowd counting.
• The meta-learning framework combined with test-time adaptation significantly improves model adaptability.
• This method provides a viable alternative to resource-intensive domain adaptation techniques.