CountSE: Soft Exemplar Open-set Object Counting

Shuai Liu, Peng Zhang, Shiwei Zhang, Wei Ke*

Official PyTorch implementation for CountSE. Please check the [Paper] for details .

CountSE Architecture

<img src=img/architecture.jpg width="100%"/>

Contents

• 📦 Preparation
• 🎯 Inference
• 🚀 Training
• 📜 Citation
• 🙏 Acknowledgements

Preparation

1. Clone Repository

git clone git@github.com:pppppz22/CountSE.git

2. Dataset Preparation

We use fsc147 to train and test our model. Click the FSC-147 to download it. Then modify the relevant paths of config/datasets_fsc147_val.json and datasets_fsc147_test. json in the root directory.

3. Set Up Environment

If you have already run CountGD, you can directly use its environment to run our code.

Install GCC and virtual environment. We used Anaconda version 2024.02-1.

sudo apt update
sudo apt install build-essential

conda create -n countse python=3.9.19
conda activate countse
cd CountSE
pip install -r requirements.txt
export CC=/usr/bin/gcc-11 # this ensures that gcc 11 is being used for compilation
cd models/GroundingDINO/ops
python setup.py build install
python test.py # should result in 6 lines of * True
cd ../../../

If you have problems during the installation environment, you can check the CountGD or Open-GrundingDINO repository issues for assistance.

4. Download Pre-Trained Weights

Download pre-training groudingdino weights

wget -P checkpoints https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha2/groundingdino_swinb_cogcoor.pth

Download Bert

python download_bert.py

Training

You can use the following command to train the model. The --output_dir argument specifies the path where the model will be saved. If you need to resume training, you can use the --start_epoch parameter, which defaults to 0.

python -u main.py --output_dir ./countse_ckpt -c config/cfg_fsc147_val.py --datasets config/datasets_fsc147_val.json --pretrain_model_path ./pretrained_ckpts/groundingdino_swinb_cogcoor.pth --gpuid 0 --options text_encoder_type=./pretrained_ckpts/bert-base-uncased

Inference

You can download our trained checkpoint to obtain the results in the paper. Update your weight path and you can run inference on a single RTX 3090 using the following command. You can switch test datasets by modifying the -c and --datasets arguments.

python -u main_inference.py --eval --output_dir ./inference_val -c config/cfg_fsc147_test.py --datasets config/datasets_fsc147_test.json --pretrain_model_path ./pretrained_ckpts/checkpoint_best_regular.pth --gpuid 0 --options text_encoder_type=./pretrained_ckpts/bert-base-uncased --crop --remove_bad_exemplar

Citation

@inproceedings{liu2025countse,
  title={CountSE: Soft Exemplar Open-set Object Counting},
  author={Liu, Shuai and Zhang, Peng and Zhang, Shiwei and Ke, Wei},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={21536--21546},
  year={2025}
}

Acknowledgements

Our code is based on CountGD. If you have any questions, please click pppzhang@stu.xjtu.edu.cn Contact me