Bayesian-Crowd-Counting （ICCV 2019 oral）

Official Implement of ICCV 2019 oral paper "Bayesian Loss for Crowd Count Estimation with Point Supervision"

Visualization

Bayesian

Bayesian+

Density

Citation

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}
}

Code

Install dependencies

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

Train and Test

1、 Dowload Dataset UCF-QNRF Link

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

python preprocess_dataset.py --origin_dir PATH_TO_ORIGIN_DATASET --data_dir PATH_TO_DATASET

python preprocess_shanghai.py --origin_dir PATH_TO_ORIGIN_DATASET --data_dir PATH_TO_DATASET --part 'A/B'

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> --dataset "qnrf/sha/shb" --max_epoch xxx --extra_aug

4、 Test Model

python test.py --data_dir <directory of processed data> --save_dir <directory of log and model> --dataset "qnrf/sha/shb"

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

Pretrain Weight

UCF-QNRF

Baidu Yun Link extract code: x9wc

Google Drive Link

ShanghaiTech A

Baidu Yun Link extract code: tx0m

Goodle Drive Link

ShanghaiTech B

Baidu Yun Link extract code: a15u

Goodle Drive Link

Reproduction

qnrf

paper: mae: 88.7, mse: 154.8

pretrained_models/best_model_sha.pth: mae 86.31, mse 153.04

best_model.pth: mae 89.61100787316968, mse 161.90780984805582

Shanghai A

paper: mae: 62.8, mse: 101.8

pretrained_models/best_model_sha.pth: mae 62.68, mse 97.30

best_model.pth: mae 63.7321413061121, mse 100.5095305210611

Shanghai B

paper: mae: 7.7, mse: 12.7

pretrained_models/best_model_shb.pth: mae 7.747125667861745, mse 12.956488955530084

best_model.pth: mae 8.229536213452302, mse 13.057458091084367

Nightmare4214/Bayesian-Crowd-Counting