[Paper Review] C^3 Framework: An Open-source PyTorch Code for Crowd Counting

Junyu Gao, Wei Lin+4 more|arXiv (Cornell University)|Jul 5, 2019

Anomaly Detection Techniques and ApplicationsComputer Science9 references35 citations

TL;DR

An open-source PyTorch code framework (C3F) for crowd counting providing solid baselines, data preprocessing strategies for major datasets, reproduction of mainstream methods, and a robust experiment logging system.

ABSTRACT

This technical report attempts to provide efficient and solid kits addressed on the field of crowd counting, which is denoted as Crowd Counting Code Framework (C$^3$F). The contributions of C$^3$F are in three folds: 1) Some solid baseline networks are presented, which have achieved the state-of-the-arts. 2) Some flexible parameter setting strategies are provided to further promote the performance. 3) A powerful log system is developed to record the experiment process, which can enhance the reproducibility of each experiment. Our code is made publicly available at \url{https://github.com/gjy3035/C-3-Framework}. Furthermore, we also post a Chinese blog\footnote{\url{https://zhuanlan.zhihu.com/p/65650998}} to describe the details and insights of crowd counting.

Motivation & Objective

• Provide a uniform, efficient code interface for crowd counting research and benchmarking.
• Offer solid baseline models derived from AlexNet, VGG, and ResNet for density map estimation.
• Deliver dataset-specific preprocessing guidelines to enable fair comparisons.
• Enable reproducibility through a comprehensive logging system and open-source tooling.
• Showcase reproduction of mainstream methods to facilitate fair benchmarking.

Proposed method

• Present solid baseline networks adapted from AlexNet, VGG, and ResNet for density map regression.
• Use encoder-decoder structures with tailored upsampling to produce 1-channel density maps.
• Propose preprocessing strategies for six datasets (UCF_CC_50, WorldExpo’10, SHTA, SHTB, UCF-QNRF, GCC) including image scale and input sizes divisible by 16.
• Discuss label transformation options including down-sampling and normalization; recommend avoiding heavy density map down-sampling due to PSNR/SSIM concerns.
• Provide a reproduction study of MCNN, CMTL, CSRNet, and SANet within the C3F framework, with practical training tricks to enhance performance.
• Offer an open-source codebase at GitHub for community use and contribution.

Experimental results

Research questions

• RQ1 Can reproducible, uniform preprocessing and code interfaces enable fair benchmarking across crowd counting methods?
• RQ2 How do standard backbones (AlexNet, VGG, ResNet) perform as density map regressors within a unified C3F framework?
• RQ3 What is the impact of dataset-specific input scaling and transformation strategies on counting accuracy (MAE/MSE) across common benchmarks?
• RQ4 To what extent can reproduction tricks align reproduced results with published performance for leading crowd counting methods?
• RQ5 Does a centralized logging system improve experiment reproducibility and traceability in crowd counting research?

Key findings

• ResNet-based models achieve strong performance with MAE 7.6–7.7 and MSE 12.2–12.6 on the tested configurations.
• VGG-based configurations yield MAE around 10.3–10.5 and MSE around 16.0–17.4, with VGG+decoder giving sharper density maps.
• Reproduced results for CSRNet closely match published figures (MAE 10.6, MSE 16.0 in original vs 10.6, 16.6 in reproduction).
• Reproductions of MCNN, CMTL, and SANet show gaps to some reported figures, but SANet reproduction comes closest to the original results (e.g., SANet original 8.4/13.6 vs 12.1/19.2 in reproduction).
• The framework provides practical data preprocessing strategies (image scale, input size divisibility by 16) that facilitate fair cross-method comparisons.
• C3F includes a robust logging system to enhance reproducibility across experiments.