One-Shot Any-Scene Crowd Counting With Local-to-Global Guidance.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Crowd counting models struggle with unseen surveillance scenes due to variations in camera installation.
• Accurate crowd counting from any scene using a single annotated image is a significant challenge.

Purpose of the Study:

• To develop a robust crowd counting method for any surveillance scene using one-shot learning.
• To enable accurate crowd estimation from limited data in diverse, unannotated environments.

Main Methods:

• Crowd counting is framed as a one-shot learning task utilizing metric learning.
• A Multi-Prototype Learner estimates foreground and density prototypes from a single support image via Expectation-Maximization.
• Local features are activated by Convolutional Neural Networks (CNNs), and global features are correlated using Transformers.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art (SOTA) methods in few-shot crowd counting scenarios.
• Experiments on three surveillance datasets validate the effectiveness of the local-to-global guidance strategy.

Conclusions:

• The developed one-shot learning approach effectively addresses the challenge of crowd counting in unseen surveillance scenes.
• The Multi-Prototype Learner and local-to-global guidance mechanism provide a robust solution for accurate crowd estimation with limited data.