fixedpoint_prompt_counting

Official code for AAAI-24 paper:"A Fixed-Point Approach to Unified Prompt-Based Counting"

The motivation is to propose a unified framework for prompt-based counting, which is able to handle both box, point and text prompts.

The pipeline of the proposed method is as follows:

Requirement

We use Singularity to build the enviroment. Download our enviroment: excalibur.sif. If you'd like to create environement yourself, the following python packages are required:

pytorch == 1.9.0
torchvision == 0.10.0
timm == 0.4.12
termcolor
yacs
einops

Data Preparation

1. Download FSC-147;
2. modify the following parameters in data_fsc147/prepare_data.sh:
   • Let ori_root be the local path of FSC-147.
   • Let new_root be the path that you want to save the modified FSC-147.
3. execute the script:

cd data_fsc147
bash prepare_data.sh

Training

• set the data_path in run.sh the same as new_root in data_fsc147/prepare_data.sh
• If you use our singularity: singularity exec --bind --nv path_to_excalibur.sif ./run.sh
• If you create the environment yourself, just execute the script: ./run.sh

A training log is shown in md-files/training.log, and corresponding checkpoint fxp.pth is uploaded here.

Citation

@inproceedings{lin2024fixed,
  title={A Fixed-Point Approach to Unified Prompt-Based Counting},
  author={Lin, Wei and Chan, Antoni B},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={4},
  pages={3468--3476},
  year={2024}
}