/HANet

Primary LanguagePython

Hybrid attention network based on progressive embedding scale-context for

crowd counting

  • The paper is under review.

  • Paper Link

Overview

avatar

Data processing

  • Shanghai Tech A randomly crop 4 patch of 128x128 from each image.
  • Shanghai Tech B randomly crop 4 patch of 256x256 from each image.
  • UCF-CC-50 randomly crop 8 patch of 128x128 from each image.
  • QNRF randomly crop 8 patch of 128x128 from each image.
  • NWPU-Crowd randomly crop 1 patch of 576x768 from each image.

Training

  • update root "HOME and DATASET" in ./config.py.
  • python train.py

Testing

  • python eval.py

Evaluation metrics

  • MSE and MSE

Visualization

  • python vis.py

Environment

python >=3.6 
pytorch >=1.5
opencv-python >=4.0
scipy >=1.4.0
h5py >=2.10
pillow >=7.0.0
imageio >=1.18