This repo is the official implementation of paper: Neuron Linear Transformation: Modeling the Domain Shift for Crowd Counting (T-NNLS, 2021). The code is developed based on C3F.

• Installation
• Project Architecture
• Training
• Test

It is recommended to prepare the following dependencies before training.

• Prerequisites
  • Python 3.7
  • Pytorch >=1.5: http://pytorch.org .
  • other libs in requirements.txt, run pip install -r requirements.txt.
• Code
  • Clone this repository in the directory (Root/NLT):

    git clone https://github.com/taohan10200/NLT.git

• Dataset downloading
  • the source dataset: GCC [Link1] [Link2] [Link3 (pwd:utdo)]
  • other target datasets: ShanghaiTech Part [Link1] [Link2]); UCF-QNRF [Homepage] [Download]; UCSD [Homepage]; MALL [Homepage]; WorldExpo'10 [Homepage].
  • Generate the density map with the scripts in Root/NLT/data_split/generate_den_map

• Finally, the folder tree is below:

   -- ProcessedData
   	|-- performed_bak_lite   # this is GCC dataset, which contains 100 scenes (a total of 400 folders).
           |-- scene_00_0
           |	   |-- pngs_544_960
           |	   |-- den_maps_k15_s4_544_960
           |-- ...
           |-- scene_99_3
           |	   |-- pngs_544_960
           |	   |-- den_maps_k15_s4_544_960
   	|-- SHHB
   	    |-- train
   	    |    |-- img
   	    |    |-- den
   	    |-- test
   	    |    |-- img
   	    |    |-- den
   	|-- ...		
   -- NLT
     |-- data_split
     |-- dataloader
     |-- models
     |-- ...

Modify the flowing configurations in config.py:

__C.model_type = 'vgg16'   # choices=['ResNet50','vgg16']
__C.phase ='pre_train'     # choices=['pre_train', 'DA_train', 'fine_tune'])
__C.gpu_id = "2,3"         # single gpu:"0"..; multi gpus:"2,3,
__C.target_dataset ='SHHB' # dataset choices =  ['SHHB',  'UCF50',  'QNRF', 'MALL', 'UCSD', 'SHHA']

Then, run the command:

python nlt_train.py

modify the __C.phase='pre_train' in config.py, and then run:

python pre_train.py

Find the pre-trained model in Root/NLT/exp/pre. Set the configurations __C.GCC_pre_train_model
and __C.phase='fine_tune' in config.py, and then run:

python pre_train.py

To evaluate the metrics (MAE\MSE\PNSR\SSIM) on test set, you should fill the model path (cc_path) and dataset name in test.py, and then run:

python test.py

The visual density map can be selectively generated in Root/NLT/visual-display.

If you find this project is useful for your research, please cite:

@article{wang2021neuron,
  title={Neuron linear transformation: modeling the domain shift for crowd counting},
  author={Wang, Qi and Han, Tao and Gao, Junyu and Yuan, Yuan},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2021},
  publisher={IEEE}
}