[Paper Review] Region-Aware Network: Model Human's Top-Down Visual Perception Mechanism for Crowd Counting

Yuehai Chen, Jing Yang+4 more|arXiv (Cornell University)|Jun 23, 2021

Video Surveillance and Tracking Methods55 references23 citations

TL;DR

This paper proposes RANet, a region-aware feedback network that models human top-down visual perception for crowd counting. By generating priority maps to highlight crowd regions and using a Region-Aware block that computes global similarity between features and priorities, the model enhances context modeling and receptive field size, achieving state-of-the-art performance on multiple crowd counting benchmarks despite challenges from background noise and scale variation.

ABSTRACT

Background noise and scale variation are common problems that have been long recognized in crowd counting. Humans glance at a crowd image and instantly know the approximate number of human and where they are through attention the crowd regions and the congestion degree of crowd regions with a global receptive field. Hence, in this paper, we propose a novel feedback network with Region-Aware block called RANet by modeling humans Top-Down visual perception mechanism. Firstly, we introduce a feedback architecture to generate priority maps that provide prior about candidate crowd regions in input images. The prior enables the RANet pay more attention to crowd regions. Then we design Region-Aware block that could adaptively encode the contextual information into input images through global receptive field. More specifically, we scan the whole input images and its priority maps in the form of column vector to obtain a relevance matrix estimating their similarity. The relevance matrix obtained would be utilized to build global relationships between pixels. Our method outperforms state-of-the-art crowd counting methods on several public datasets.

Motivation & Objective

• To address background noise and scale variation in crowd counting, which hinder existing deep learning methods.
• To model human-like top-down visual perception, where attention is guided by prior knowledge of crowd regions.
• To improve feature representation by expanding the effective receptive field using global context information.
• To achieve state-of-the-art performance on standard crowd counting benchmarks.

Proposed method

• Proposes a feedback architecture that generates a priority map to indicate likely crowd regions, reducing background interference.
• Introduces a Region-Aware block that computes a relevance matrix by measuring similarity between flattened input image and priority map as column vectors.
• Uses the relevance matrix to reweight features, encoding global context and enhancing relationships across distant pixels.
• Employs a global receptive field mechanism to better handle scale variation in dense crowd scenes.
• Combines attention-based feature refinement with global context aggregation to improve density estimation.
• Trains end-to-end with a standard regression loss for density map prediction.

Experimental results

Research questions

• RQ1 Can modeling human top-down visual perception improve crowd counting accuracy in complex scenes?
• RQ2 How can global context and long-range dependencies be effectively modeled in crowd counting networks?
• RQ3 Can a feedback mechanism that generates priority maps enhance attention to crowd regions and suppress background noise?
• RQ4 To what extent does expanding the effective receptive field improve performance on scale-variant crowd scenes?
• RQ5 Does integrating global context through a similarity-based mechanism outperform local or pixel-wise attention in crowd counting?

Key findings

• RANet achieves state-of-the-art performance on multiple public crowd counting datasets, including UCF-QNRF, ShanghaiTech, and UCSD.
• The proposed feedback network with priority maps significantly reduces attention to background clutter, improving robustness.
• The Region-Aware block effectively expands the effective receptive field by modeling global relationships, enhancing scale generalization.
• Quantitative results show consistent improvements in MAE and MSE across all benchmarks, with lower error rates than prior SOTA methods.
• Ablation studies confirm that both the priority map generation and global context modeling components are essential for performance gains.
• The method demonstrates strong generalization on highly congested scenes, where scale variation and occlusion are most severe.