DCL-CrowdCounting

This is an official implementaion of the paper "Density-aware Curriculum Learning for Crowd Counting", completed in November 2019, accepted by T-CYB in October 2020.

[IEEE link][pdf download]

This repository shows how PSCC is trained with/without DCL strategy. Relevant experiment processes are shown in process_reports.

• normal.log demonstrates the process of PSCC under random sampling.
• curriculum.log demonstrates the process of PSCC under density-aware curriculum learning.
• *.txt shows the configration and verification results during training.

Requirements

• Python 2.7 (It is 2019 when submiting the paper. py3 will be supported in the future.)
• Pytorch 1.2.0
• TensorboardX
• torchvision 0.4.0
• easydict

Dara preparation

1. Download the original ShanghaiTech Dataset [link: Dropbox / BaiduPan]
2. generate the density maps using the datasets/generate_data.py (using Python 3 because of the f-string) according to the README in datasets.
3. modify the dataset/SHHA/setting.py th specify the path of dataset.

Training

1. modify the training parameters in config.py.
   • Without DCL, set __C.DCL_CONF['work'] = False
   • With DCL, set __C.DCL_CONF['work'] = True
2. python train.py

Experiment Results

PSCC	MAE	MSE
Random Sampling	66.82	109.35
Density-aware CL	64.97	107.96

Citation

If you use the code, please cite the following paper:

@ARTICLE{9275392,
  author={Q. {Wang} and W. {Lin} and J. {Gao} and X. {Li}},
  journal={IEEE Transactions on Cybernetics}, 
  title={Density-Aware Curriculum Learning for Crowd Counting}, 
  year={2020},
  volume={},
  number={},
  pages={1-13},
  doi={10.1109/TCYB.2020.3033428}}