JHU-CROWD++: Large-Scale Crowd Counting Dataset and A Benchmark Method.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Crowd counting is crucial for public safety and resource management.
• Existing datasets often lack diversity in scenarios, weather, and illumination, limiting model generalizability.
• Accurate crowd density estimation in unconstrained environments remains a significant challenge.

Purpose of the Study:

• To introduce JHU-CROWD++, a large-scale, diverse dataset for unconstrained crowd counting.
• To propose a novel deep learning network for progressive crowd density map generation.
• To evaluate the effectiveness of the proposed network on challenging datasets.

Main Methods:

• Developed JHU-CROWD++ dataset with 4,372 images and 1.51 million annotations, featuring diverse conditions (weather, illumination).
• Proposed Confidence Guided Deep Residual Counting Network (CG-DRCN) using VGG16 backbone.
• Employed progressive density map generation via residual error estimation with uncertainty-based confidence weighting.

Main Results:

• JHU-CROWD++ dataset provides a challenging benchmark for crowd counting research.
• CG-DRCN achieved significant improvements in crowd density estimation errors compared to recent methods.
• The confidence weighting mechanism effectively guided the refinement of density maps.

Conclusions:

• JHU-CROWD++ dataset advances crowd counting research by offering unprecedented diversity and scale.
• CG-DRCN demonstrates a promising approach for accurate crowd density estimation in complex scenarios.
• The proposed method highlights the importance of uncertainty-aware learning for robust crowd counting.