Density-Aware Curriculum Learning for Crowd Counting.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Crowd counting is vital for public safety and congestion management.
• Deep learning models for crowd counting are effective but computationally expensive and time-consuming to train.
• Existing models often struggle with scale and complexity.

Purpose of the Study:

• To develop a lightweight yet effective crowd counting model.
• To introduce an efficient training strategy to enhance model performance and generalization.
• To address the limitations of large, complex deep learning models in crowd counting.

Main Methods:

• A novel lightweight model, the Pixel Shuffle Decoder (PSD), was constructed, featuring an image feature encoder and a pixel shuffle operator for enhanced density information.
• A Density-Aware Curriculum Learning (DCL) strategy was designed, assigning weights to pixels based on prediction difficulty to guide training.
• The PSD model was trained using the DCL strategy, and DCL was also applied to existing crowd counting models.

Main Results:

• The PSD model achieved outstanding performance on mainstream crowd counting datasets when trained with DCL.
• Applying DCL to existing crowd counters resulted in significant performance improvements.
• The proposed methods demonstrate effectiveness in improving crowd counting accuracy and training efficiency.

Conclusions:

• The Pixel Shuffle Decoder (PSD) offers an efficient and effective solution for crowd counting.
• Density-Aware Curriculum Learning (DCL) is a versatile training strategy that enhances the performance of various crowd counting models.
• These advancements contribute to more efficient and accurate crowd density estimation for practical applications.