gaobb / CSTrans

CSTrans: Correlation-guided Self-Activation transformer for counting everything

CSTrans for Counting Everything

This is the official implementation of CSTrans. This repo is created by Zhongyi Huang.

CSTrans Framework

Visualization Results

Preparatory work

The dataset can be downloaded from here: https://drive.google.com/file/d/1ymDYrGs9DSRicfZbSCDiOu0ikGDh5k6S/view?usp=sharing.

The precomputed density maps can be found here: https://archive.org/details/FSC147-GT.

The whole project directory structure is:

PROJECT_DIR
  |--datasets
       |--FSC-147
            |--annotation_FSC-147.json
            |--gt_density_map_adaptive
            |--gt_density_map_adaptive_vis
            |--ImageClasses_FSC-147.txt
            |--img
            |--img_vis
            |--Train_Test_Val_FSC-147.json
  |--cst (the folder name of this source codes)
  |--outputs (auto-created when running this source codes)
  |--torchvision_pretrained_models

Before running the codes, please specify the PROJECT_DIR in tools/constants.py.

Training

The training commands are summarized in the run_train.sh.

Testing

For testing, please set the dir_list_list of run_test.py, and run:

python3 run_test.py

Additional Information

If you find CSTrans helpful, please cite it as

@article{gao2024cstrans,
  title={CSTrans: Correlation-guided Self-Activation transformer for counting everything},
  author={Gao, Bin-Bin and Huang, Zhongyi},
  journal={Pattern Recognition},
  pages={110556},
  year={2024},
  publisher={Elsevier}
}

Acknowledgement

This repo is developed based on FamNet.