/CSRNet-pytorch

The simple implemention of CSRNet.

Primary LanguagePythonMIT LicenseMIT

CSRNet-Simple-Pytorch

This is an simple and clean implemention of CVPR 2018 paper "CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes".

Requirement

  1. Install pytorch 1.0.0+, tensorboardX, opencv, etc.

    conda remove --name pytorch --all
conda create -n pytorch -y python=3.7.3 numpy scipy ipykernel
conda activate pytorch
conda install -y -c conda-forge opencv matplotlib h5py tensorboardx tqdm pylint
conda install -y -c pytorch pytorch=1.4.0 torchvision=0.5.0 cudatoolkit=10.1
ipython kernel install --user --name=pytorch
conda deactivate

  2. Clone this repository

    git clone https://github.com/CommissarMa/CSRNet-pytorch.git

    We'll call the directory that you cloned CSRNet-pytorch as ROOT.

Data Setup

  1. Download ShanghaiTech Dataset from Dropbox or Baidu Disk (code: a2v8)
  2. Put ShanghaiTech Dataset in 'ROOT/data'. You can find two python scripts in 'data_preparation' folder which are used to generate ground truth density-map for ShanghaiTech PartA and PartB respectively. (Mind that you need move the script to corresponding sub dataset folder like 'ROOT/data/part_A_final' and run it)

Train

  1. Modify the dataset root in 'config.py'
  2. Run 'train.py'
  3. Open the command line and type in 'tensorboard --logdir=ROOT/runs', then browse 'localhost:6006' to see the visual result.

Testing

  1. Run 'test.py' for calculate MAE of test images or just show an estimated density-map.

Other notes

  1. We trained the model and got 67.74 MAE at 124-th epoch on ShanghaiTech PartA. Got 8.68 MAE on PartB at 94-th epoch.
  2. If you are new to crowd counting, we recommand you to take a look at Crowd_counting_from_scratch first. It is a tutorial for crowd counting beginner.

Download Crowd Data