Home

Awesome

FusionCount

Official PyTorch implementation of the ICIP 2022 paper FusionCount: Efficient Crowd Counting via Multiscale Feature Fusion.

Requirements

The code was writen in Python 3.9.10. The PyTorch version used is 1.10.0

Datasets

ShanghaiTech A & B can be found on Kaggle.

Model Structure

FusionCount

Training

Please refer to DM-Count for data preparation and training details. Please notice that you will also have to normalise the predicted density map, as illustrated in the code snippet from DM-Count/models.py below.

...
mu = self.density_layer(x)
B, C, H, W = mu.size()
mu_sum = mu.view([B, -1]).sum(1).unsqueeze(1).unsqueeze(2).unsqueeze(3)
mu_normed = mu / (mu_sum + 1e-6)
return mu, mu_normed

Results

ModelMulti-AddsMAE (SHA)RMSE (SHA)MAE (SHB)RMSE (SHB)
CSRNet856.99 G68.2115.010.616.0
CAN908.05 G62.3100.07.812.2
BL853.70 G62.8101.87.712.7
DM-Count853.70 G59.795.77.411.8
FusionCount (ours)815.00 G62.2101.26.911.8

sha_img_96 sha_img_116 shb_img_21 shb_img_30

Citation

The paper has been accepted by ICIP 2022 and published on IEEE Xplore. You can also find the arXiv version here. Please cite us if you find it useful!

@INPROCEEDINGS{9897322,
  author={Ma, Yiming and Sanchez, Victor and Guha, Tanaya},
  booktitle={2022 IEEE International Conference on Image Processing (ICIP)},
  title={Fusioncount: Efficient Crowd Counting Via Multiscale Feature Fusion},
  year={2022},
  volume={},
  number={},
  pages={3256-3260},
  doi={10.1109/ICIP46576.2022.9897322}
}