CDSpixel

The official implementation for aaai24 Learning Invariant Inter-pixel Correlations for Superpixel Generation'

This repo is based on SCN. So the installation and data preparation is pretty similar.

Installation

Step 0. Install PyTorch and Torchvision following official instructions, e.g.,

pip install torch torchvision
# The training code was mainly developed and tested with python 3.7, PyTorch 1.8, CUDA 10

Step 1. Follow the protocol of SCN and SSN, compile the component connection method in SSN to enforce the connectivity in superpixels.

cd third_party/cython/
python setup.py install --user
cd ../..

Dataset Preparation

BSDS500

Download and convert the primary training dataset, i.e., BSDS, and put it in <BSDS_DIR> folder.

mkdir -p data/BSDS
# Put the downloaded archive in this folder

python ./tools/pre_process_bsd500.py --dataset=<BSDS_DIR> --dump_root=<DUMP_DIR>
# E.g., python ./tools/pre_process_bsd500.py --dataset='./data/BSDS' --dump_root='./data/BSDS'

python pre_process_bsd500_ori_sz.py --dataset=<BSDS_DIR> --dump_root=<DUMP_DIR>
# E.g., python ./tools/pre_process_bsd500_ori_sz.py --dataset='./data/BSDS' --dump_root='./data/BSDS'

The code will generate three folders under the <DUMP_DIR>, named as /train, /val, and /test, and three .txt files record the absolute path of the images, named as train.txt, val.txt, and test.txt.

Custom Datasets

we evaluate our methods mainly on NYUv2, VOC2012, KITTI. Since training or evaluation requires csv labels, use the following command to convert the label format, e.g.,

python ./tools/trans_label.py --dataset='VOC2012' --label_path=<LABEL_DIR> --trans_path=<CSV_DIR>

<LABEL_DIR> is the segmentation label path.

Citation

If you find this project useful in your research, please consider cite:

@inproceedings{xu2024leaning,
  title={Learning Invariant Inter-pixel Correlations for Superpixel Generation},
  author={Xu, Sen and Wei, Shikui and Ruan, Tao and Liao, Lixin},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={6},
  pages={6351-6359},
  year={2024},
  DOI={10.1609/aaai.v38i6.28454},
}

cdspixel's People