Learning Crowd Scale and Distribution for Weakly Supervised Crowd Counting and Localization (TCSVT)

Introduction

This is the official PyTorch implementation of paper: Learning Crowd Scale and Distribution for Weakly Supervised Crowd Counting and Localization (extended from paper Weakly-supervised scene-specific crowd counting using real-synthetic hybrid data). This paper proposes a weakly supervised crowd counting and localization method based on scene-specific synthetic data for surveillance scenarios, which can accurately predict the number and location of person without any manually labeled point-wise or count-wise annotations.

Getting started

preparatoin

• Clone this repo in the directory

• Install dependencies. We use python 3.7 and pytorch == 1.10.0 : http://pytorch.org.

  conda create -n LCSD python=3.7
  conda activate LCSD
  conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge
  cd ${LCSD}
  pip install -r requirements.txt
  pip install git+http://github.com/nikitadurasov/masksembles
  

• Download the resource files used by the code from this link, including datasets, pre-trained models, pedestrian gallery, and negetive samples. Unzip resources.zip to resources. The resources folder is organized as follows:

  $resources/
  ├── CityUHK-X  # dataset name
  │   ├── scene_001  # scene name
  │   │   ├── CityUHK-X_scene_001_20_40  # specific dataset for this scene
  │   │   │   ├── train_data
  │   │   │   │   ├── images
  │   │   │   │   │   └── xx.jpg
  │   │   │   │   ├── ground_truth_txt
  │   │   │   │   │   └── xx.txt
  │   │   │   ├── test_data
  │   │   │   ├── train_data.txt
  │   │   │   └── test_data.txt
  │   │   └── scene.jpg  #  scene image without person
  │   ├── scene_002
  │   ├── ...
  │   └── scene_k
  ├── Mall
  │   ├── scene_001  # only one scene for Mall
  │   │   ├── mall_800_1200
  │   │   │   ├── train_data
  │   │   │   │   ├── images
  │   │   │   │   │   └── xx.jpg
  │   │   │   │   ├── ground_truth_txt
  │   │   │   │   │   └── xx.txt
  │   │   │   ├── test_data
  │   │   │   ├── train_data.txt
  │   │   │   └── test_data.txt
  │   │   └── scene.jpg
  ├── UCSD
  │   ├── scene_001
  │   │   ├── ucsd_800_1200
  │   │   │   ├── train_data
  │   │   │   │   ├── images
  │   │   │   │   │   └── xx.jpg
  │   │   │   │   ├── ground_truth_txt
  │   │   │   │   │   └── xx.txt
  │   │   │   ├── test_data
  │   │   │   ├── train_data.txt
  │   │   │   └── test_data.txt
  │   │   └── scene.jpg
  ├── pedestrians  #  pedestrian gallery
  │   ├── GCC #  default
  │   │   └── xx.png
  │   ├── SHHB
  │   └── LSTN
  ├── indoor_negetive_samples 
  │   └── xx.jpg
  ├── outdoor_negetive_samples
  │   └── xx.jpg
  ├── darknet53.conv.74  #  pre-trained model for detection
  └── net_G_last.pth.txt  #  pre-trained model for image harmonization
  

Training

Check some parameters in train.py before training:

• Use dataset = Mall to set the dataset.
• Use scene = scene_001 to set the scene of the dataset. Mall and UCSD only have one scene, so set scene as scene_001.
• Use resource-path = {$resources} to set the path of the resource folder downloaded above.
• Use scene-dataset = mall_800_1200 to set the specific dataset of the scene.
• Use device = 0 to set the gpu id for training.
• run python train.py.

Test

Check some parameters in test.py before test:

• Use dataset = Mall to set the dataset.
• Use scene = scene_001 to set the scene of the dataset.
• Use resource-path = {$resources} to set the path of the resource folder.
• Use scene-dataset = mall_800_1200 to set the specific dataset of the scene.
• Use model_path = xxx.pth to set the pre-trained counter in the training stage.
• Use test-name = xxx to set the test name, which will be used to name the folder for saving the test results.
• Use device = 0 to set the gpu used for test.
• run test.py

Citation

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

@ARTICLE{LCSD,
  author={Fan, Yaowu and Wan, Jia and Ma, Andy J.},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Learning Crowd Scale and Distribution for Weakly Supervised Crowd Counting and Localization}, 
  year={2025},
  volume={35},
  number={1},
  pages={713-727}
  }

@INPROCEEDINGS{ICASSP_2023_FAN
  author={Fan, Yaowu and Wan, Jia and Yuan, Yuan and Wang, Qi},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
  title={Weakly-Supervised Scene-Specific Crowd Counting Using Real-Synthetic Hybrid Data}, 
  year={2023},
  pages={1-5}
}

Acknowledgement

The released PyTorch training script borrows some codes from the masksembles, yolov3, and RainNet. If you think this repo is helpful for your research, please consider cite them.