NWPU-Crowd Sample Code

This repo is the official implementation of paper: NWPU-Crowd: A Large-Scale Benchmark for Crowd Counting. The code is developed based on C^3 Framework.

Compared with the original C^3 Framework,

• the python3.x's new features are utilized;
• the density map is generated online by a conv layer for saving io time on the disk;
• improve the visualization in the Tensorboard.

These features will be merged into C^3 Framework as soon as possible.

Getting Started

Preparation

• Prerequisites

  • Python 3.x
  • Pytorch 1.x: http://pytorch.org .
  • other libs in requirements.txt, run pip install -r requirements.txt.

• Installation

  • Clone this repo:

    git clone https://github.com/gjy3035/NWPU-Crowd-Sample-Code.git

• Data Preparation

  • Download NWPU-Crowd dataset from OneDrive1, OneDrive2 or BaiduNetDisk.
  • Unzip *zip files in turns and place images_part* into a folder. Finally, the folder tree is below:

    -- NWPU-Crowd
    |-- images
    |   |-- 0001.jpg
    |   |-- 0002.jpg
    |   |-- ...
    |   |-- 5109.jpg
    |-- jsons
    |   |-- 0001.json
    |   |-- 0002.json
    |   |-- ...
    |   |-- 3609.json
    |-- mats
    |   |-- 0001.mat
    |   |-- 0002.mat
    |   |-- ...
    |   |-- 3609.mat
    |-- train.txt
    |-- val.txt
    |-- test.txt
    |-- readme.md

  • Run ./datasets/prepare_NWPU.m using Matlab.
  • Modify __C_NWPU.DATA_PATH in ./datasets/setting/NWPU.py with the path of your processed data.

Training

• Set the parameters in config.py and ./datasets/setting/NWPU.py (if you want to reproduce our results, you are recommended to use our parameters in ./saved_exp_para).
• run python train.py.
• run tensorboard --logdir=exp --port=6006.

Testing

We only provide an example to forward the model on the test set. You may need to modify it to test your models.

• Modify some key parameters in test.py:
  • Line 32: LOG_PARA, the same as __C_NWPU.LOG_PARA in ./datasets/setting/NWPU.py.
  • Line 34: dataRoot, the same as __C_NWPU.DATA_PATH in ./datasets/setting/NWPU.py.
  • Line 36: model_path.
  • Line 48: GPU Id and Model Name.
• Run python test.py.

Pre-trained Models

We provide the pre-trained models in this link, which is a temporary share point of OneDrive. We will provide a permanent website ASAP.

Performance on the validation set

For an intuitive comparison, the visualization results of these methods are provided at this link. The overall results on val set:

About the leaderboard on the test set, please visit Crowd benchmark.

References

1. Single-Image Crowd Counting via Multi-Column Convolutional Neural Network, CPVR, 2016.
2. C^3 Framework: An Open-source PyTorch Code for Crowd Counting, arXiv, 2019.
3. CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes, CVPR, 2018.
4. Context-Aware Crowd Counting, CVPR, 2019.
5. SCAR: Spatial-/Channel-wise Attention Regression Networks for Crowd Counting, Neurocomputing, 2019.
6. Learning from Synthetic Data for Crowd Counting in the Wild, CVPR, 2019.

Evaluation Scheme

The Evaluation Python Code of the crowdbenchmark.com is shown in ./misc/evaluation_code.py, which is similar to our validation code in trainer.py.

Citation

If you find this project is useful for your research, please cite:

@article{gao2020nwpu,
  title={NWPU-Crowd: A Large-Scale Benchmark for Crowd Counting and Localization},
  author={Wang, Qi and Gao, Junyu and Lin, Wei and Li, Xuelong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  doi={10.1109/TPAMI.2020.3013269},
  year={2020}
}

Our code borrows a lot from the C^3 Framework, you may cite:

@article{gao2019c,
  title={C$^3$ Framework: An Open-source PyTorch Code for Crowd Counting},
  author={Gao, Junyu and Lin, Wei and Zhao, Bin and Wang, Dong and Gao, Chenyu and Wen, Jun},
  journal={arXiv preprint arXiv:1907.02724},
  year={2019}
}

If you use crowd counting models in this repo (MCNN, C3F-VGG, CSRNet, CANNet, SCAR, and SFCN+), please cite them.