Kernel-Based Density Map Generation for Dense Object Counting.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Crowd counting is vital for surveillance, typically using two-step density map generation and estimation.
• Current methods use hand-crafted density map generation, which may not be optimal for end-to-end training.
• Existing density map-based methods have significantly improved counting performance.

Purpose of the Study:

• To develop an adaptive density map generator for improved end-to-end crowd counting.
• To create a learnable density map representation for object counting tasks.
• To enhance the overall accuracy and adaptability of crowd counting algorithms.

Main Methods:

• Proposed an adaptive density map generator taking annotation dot maps as input.
• Developed a joint end-to-end training framework for the generator and counter.
• Applied the framework to crowd counting, vehicle counting, and general object counting.

Main Results:

• Demonstrated the effectiveness of learnable density map representations.
• Achieved improved counting performance through the proposed adaptive generator.
• Validated the framework's applicability and success across 10 diverse datasets.

Conclusions:

• The proposed adaptive density map generator significantly enhances crowd counting and general object counting.
• End-to-end training with learnable density maps offers superior performance over traditional methods.
• The framework provides a flexible and effective solution for dense object counting applications.