Deep Rank-Consistent Pyramid Model for Enhanced Crowd Counting

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Conventional crowd counting relies on fully-supervised learning, requiring extensive pixel-level annotations.
• Labeling is costly and time-intensive, motivating the use of unlabeled data.
• Unlabeled images offer inherent structural information and rank consistency for supervision.

Purpose of the Study:

• To develop a novel crowd counting method that effectively utilizes unlabeled images.
• To reduce the reliance on costly pixel-level annotations in crowd counting models.
• To enhance crowd counting accuracy by leveraging rank consistency in latent feature spaces.

Main Methods:

• Proposes the Deep Rank-consistent pyramid Model (DREAM).
• Utilizes rank consistency within latent feature spaces across coarse-to-fine pyramid features.
• Incorporates numerous pyramid partial orders for stronger model representation.
• Introduces a new unlabeled crowd counting dataset (FUDAN-UCC) with 4000 images.

Main Results:

• Demonstrates the effectiveness of DREAM on benchmark datasets (UCF-QNRF, ShanghaiTech Part A/B, UCF-CC-50).
• Achieves improved performance compared to previous semi-supervised crowd counting methods.
• Shows increased utilization of unlabeled samples.

Conclusions:

• DREAM effectively leverages rank consistency in latent feature spaces for crowd counting.
• The proposed method significantly reduces the need for manual annotations.
• DREAM offers a promising approach for crowd counting using large amounts of unlabeled data.