Arxiv | CVF

If you use this code for your research, please cite our paper:

@inproceedings{ma2019bayesian,
  title={Bayesian loss for crowd count estimation with point supervision},
  author={Ma, Zhiheng and Wei, Xing and Hong, Xiaopeng and Gong, Yihong},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={6142--6151},
  year={2019}
}

torch >= 1.0 torchvision opencv numpy scipy, all the dependencies can be easily installed by pip or conda

This code was tested with python 3.6

1、 Dowload Dataset UCF-QNRF Link

2、 Pre-Process Data (resize image and split train/validation)

python preprocess_dataset.py --origin_dir <directory of original data> --data_dir <directory of processed data>

3、 Train model (validate on single GTX Titan X)

python train.py --data_dir <directory of processed data> --save_dir <directory of log and model>

4、 Test Model

python test.py --data_dir <directory of processed data> --save_dir <directory of log and model>

The result is slightly influenced by the random seed, but fixing the random seed (have to set cuda_benchmark to False) will make training time extrodinary long, so sometimes you can get a slightly worse result than the reported result, but most of time you can get a better result than the reported one. If you find this code is useful, please give us a star and cite our paper, have fun.

5、 Training on ShanghaiTech Dataset

Change dataloader to crowd_sh.py

For shanghaitech a, you should set learning rate to 1e-6, and bg_ratio to 0.1

Baidu Yun Link extract code: x9wc

Google Drive Link

Baidu Yun Link extract code: tx0m

Goodle Drive Link

Baidu Yun Link extract code: a15u

Goodle Drive Link

GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007 Copyright © 2007 Free Software Foundation, Inc. http://fsf.org/