Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RepSFNet : A Single Fusion Network with Structural Reparameterization for Crowd Counting

Mas Nurul Achmadiah, Chi-Chia Sun, Wen-Kai Kuo, Jun-Wei Hsieh

TL;DR

RepSFNet tackles real-time crowd counting in scenes with extreme density variation by combining a RepLK-ViT backbone with reparameterized large kernels, a multi-scale Feature Fusion of ASPP and CAN, and a Concatenate Fusion to preserve spatial detail. The model is trained with a joint objective that blends Mean Absolute Error and Optimal Transport loss, enabling both accurate counts and spatially faithful density maps. Empirical results on ShanghaiTech, NWPU, and UCF-QNRF show strong performance with up to 34% lower latency than baselines, highlighting suitability for edge and low-power deployments. The approach emphasizes efficiency through structured reparameterization and fusion design, while noting limitations in highly occluded or sparse scenarios and proposing future work on lightweight attention and adaptive dilation.

Abstract

Crowd counting remains challenging in variable-density scenes due to scale variations, occlusions, and the high computational cost of existing models. To address these issues, we propose RepSFNet (Reparameterized Single Fusion Network), a lightweight architecture designed for accurate and real-time crowd estimation. RepSFNet leverages a RepLK-ViT backbone with large reparameterized kernels for efficient multi-scale feature extraction. It further integrates a Feature Fusion module combining Atrous Spatial Pyramid Pooling (ASPP) and Context-Aware Network (CAN) to achieve robust, density-adaptive context modeling. A Concatenate Fusion module is employed to preserve spatial resolution and generate high-quality density maps. By avoiding attention mechanisms and multi-branch designs, RepSFNet significantly reduces parameters and computational complexity. The training objective combines Mean Squared Error and Optimal Transport loss to improve both count accuracy and spatial distribution alignment. Experiments conducted on ShanghaiTech, NWPU, and UCF-QNRF datasets demonstrate that RepSFNet achieves competitive accuracy while reducing inference latency by up to 34 percent compared to recent state-of-the-art methods, making it suitable for real-time and low-power edge computing applications.

RepSFNet : A Single Fusion Network with Structural Reparameterization for Crowd Counting

TL;DR

RepSFNet tackles real-time crowd counting in scenes with extreme density variation by combining a RepLK-ViT backbone with reparameterized large kernels, a multi-scale Feature Fusion of ASPP and CAN, and a Concatenate Fusion to preserve spatial detail. The model is trained with a joint objective that blends Mean Absolute Error and Optimal Transport loss, enabling both accurate counts and spatially faithful density maps. Empirical results on ShanghaiTech, NWPU, and UCF-QNRF show strong performance with up to 34% lower latency than baselines, highlighting suitability for edge and low-power deployments. The approach emphasizes efficiency through structured reparameterization and fusion design, while noting limitations in highly occluded or sparse scenarios and proposing future work on lightweight attention and adaptive dilation.

Abstract

Crowd counting remains challenging in variable-density scenes due to scale variations, occlusions, and the high computational cost of existing models. To address these issues, we propose RepSFNet (Reparameterized Single Fusion Network), a lightweight architecture designed for accurate and real-time crowd estimation. RepSFNet leverages a RepLK-ViT backbone with large reparameterized kernels for efficient multi-scale feature extraction. It further integrates a Feature Fusion module combining Atrous Spatial Pyramid Pooling (ASPP) and Context-Aware Network (CAN) to achieve robust, density-adaptive context modeling. A Concatenate Fusion module is employed to preserve spatial resolution and generate high-quality density maps. By avoiding attention mechanisms and multi-branch designs, RepSFNet significantly reduces parameters and computational complexity. The training objective combines Mean Squared Error and Optimal Transport loss to improve both count accuracy and spatial distribution alignment. Experiments conducted on ShanghaiTech, NWPU, and UCF-QNRF datasets demonstrate that RepSFNet achieves competitive accuracy while reducing inference latency by up to 34 percent compared to recent state-of-the-art methods, making it suitable for real-time and low-power edge computing applications.
Paper Structure (14 sections, 4 equations, 3 figures, 6 tables)

This paper contains 14 sections, 4 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Large Kernel Design in CNNs
  4. Density Map Ground Truth
  5. Method
  6. Reparameterizing Large Kernels
  7. Feature Fusion and Concatenate Fusion
  8. RepSFNet (Reparameterize Single Fusion Network)
  9. Mean Absolute Error (MAE) and Optimal Transport Loss Function
  10. Experimental Evaluation
  11. Experimental results and discussion
  12. Performance Comparison
  13. Ablation Study
  14. Conclusions

Figures (3)

  • Figure 1: A Simple Single Fusion Network Architecture
  • Figure 2: The complete architecture of RepSFNet
  • Figure 3: Atrous Spatial Pyramid Pooling