This is the official PyTorch implementation of paper: DR.VIC: Decomposition and Reasoning for Video Individual Counting. Different from the single image counting methods, it counts the total number of the pedestrians in a video sequence with a person in different frames only being calculated once. DRNet decomposes this new task to estimate the initial crowd number in the first frame and integrate differential crowd numbers in a set of following image pairs (namely current frame and preceding frame).

• Testing Code (2022.3.19)
• PyTorch pretrained models (2022.3.19)
• Training Code
  • HT21
  • SenseCrowd (2022.9.30)

• Clone this repo in the directory (Root/DRNet):

• Install dependencies. We use python 3.7 and pytorch >= 1.6.0 : http://pytorch.org.

  conda create -n DRNet python=3.7
  conda activate DRNet
  conda install pytorch==1.7.0 torchvision==0.8.0 cudatoolkit=10.2 -c pytorch
  cd ${DRNet}
  pip install -r requirements.txt

• PreciseRoIPooling for extracting the feature descriptors

  Note: the PreciseRoIPooling [1] module is included in the repo, but it's likely to have some problems when running the code:

  1. If you are prompted to install ninja, the following commands will help you.

     wget https://github.com/ninja-build/ninja/releases/download/v1.8.2/ninja-linux.zip
     sudo unzip ninja-linux.zip -d /usr/local/bin/
     sudo update-alternatives --install /usr/bin/ninja ninja /usr/local/bin/ninja 1 --force 

  2. If you encounter errors when compiling the PreciseRoIPooling, you can look up the original repo's issues for help. One solution to the most common errors can be found in this blog.

• Datasets

  • HT21 dataset: Download CroHD dataset from this link. Unzip HT21.zip and place HT21 into the folder (Root/dataset/).
  • SenseCrowd dataset: Download the dataset from Baidu disk or from the original dataset link.
  • Download the lists of train/val/test sets at link1 or link2, and place them to each dataset folder, respectively.

Check some parameters in config.py before training,

• Use __C.DATASET = 'HT21' to set the dataset (default: HT21).
• Use __C.GPU_ID = '0' to set the GPU.
• Use __C.MAX_EPOCH = 20 to set the number of the training epochs (default:20).
• Use __C.EXP_PATH = os.path.join('./exp', __C.DATASET) to set the dictionary for saving the code, weights, and resume point.

Check other parameters (TRAIN_BATCH_SIZE, TRAIN_SIZE etc.) in the Root/DRNet/datasets/setting in case your GPU's memory is not support for the default setting.

• run python train.py.

Tips: The training process takes ~10 hours on HT21 dataset with one TITAN RTX (24GB Memory).

To reproduce the performance, download the pre-trained models from onedrive or badu disk and then place pretrained_models folder to Root/DRNet/model/

• for HT21:
  • Run python test_HT21.py.
• for SenseCrowd:
  • Run python test_SENSE.py. Then the output file (*_SENSE_cnt.py) will be generated.

The results on HT21 and SenseCrowd.

• HT21 dataset

• SenseCrowd dataset

Please visit bilibili or YouTube to watch the video demonstration.

1. Acquisition of Localization Confidence for Accurate Object Detection, ECCV, 2018.
2. Very Deep Convolutional Networks for Large-scale Image Recognition, arXiv, 2014.
3. Feature Pyramid Networks for Object Detection, CVPR, 2017.

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

@article{han2022drvic,
  title={DR.VIC: Decomposition and Reasoning for Video Individual Counting},
  author={Han, Tao, Bai Lei, Gao, Junyu, Qi Wang, and Ouyang  Wanli},
  booktitle={CVPR},
  year={2022}
}

The released PyTorch training script borrows some codes from the C^3 Framework and SuperGlue repositories. If you think this repo is helpful for your research, please consider cite them.