Redesigning Multi-Scale Neural Network for Crowd Counting.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Crowd counting is challenging due to perspective distortions and crowd variations.
• Existing multi-scale deep neural networks (DNNs) struggle with per-pixel performance discrepancies when merging density maps.

Purpose of the Study:

• To redesign multi-scale DNNs for improved crowd counting accuracy.
• To address the limitations of current methods in merging multi-scale density maps.

Main Methods:

• Introduced a hierarchical mixture of density experts for merging multi-scale density maps.
• Implemented an expert competition and collaboration scheme with pixel-wise soft gating nets.
• Developed a novel relative local counting loss, complementary to absolute error loss.

Main Results:

• Achieved state-of-the-art performance on five public crowd counting datasets.
• Demonstrated the effectiveness of the hierarchical merging and new loss function.
• The proposed method significantly improves crowd density estimation.

Conclusions:

• The redesigned multi-scale network with hierarchical expert merging and relative local counting loss offers superior crowd counting performance.
• This work provides a more sophisticated approach to combining multi-scale information in DNNs for crowd counting.
• The method sets a new benchmark for accuracy in crowd counting tasks.