Attention-Guided Collaborative Counting.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Existing crowd counting methods often use multi-branch designs to handle scale diversity.
• These branches typically operate competitively, limiting overall performance.
• There is a need for collaborative mechanisms to enhance crowd counting accuracy.

Purpose of the Study:

• To propose a novel attention-guided collaborative counting module (AGCCM) for improved crowd counting.
• To enhance collaboration between different network branches in crowd counting models.
• To address the scale diversity problem in crowd counting more effectively.

Main Methods:

• Developed an Attention-Guided Collaborative Counting Module (AGCCM) with an Attention-Guided Module (AGM) and a Collaborative Counting Module (CCM).
• CCM recombines branch outputs for independent and joint counts, promoting collaboration.
• AGM uses a transformer with foreground-background loss functions to leverage branch advantages without extra labels.
• Introduced bidirectional transformers (Bi-Transformers) to decouple global attention into row and column attention, reducing complexity and enabling arbitrary resolution image processing.

Main Results:

• The proposed AGCCM significantly enhances collaboration among network branches.
• The AGM effectively distinguishes branch advantages using attention maps and specialized loss functions.
• Bi-Transformers reduce computational load and allow processing of full-resolution images.
• Experiments on public datasets show superior performance compared to state-of-the-art crowd counting methods.

Conclusions:

• The AGCCM represents a significant advancement in crowd counting by promoting inter-branch collaboration.
• The method effectively handles scale diversity and complex crowd scenes.
• The approach offers computational efficiency and flexibility in handling various image resolutions, outperforming existing techniques.