README.md

metadata

license: mit
task_categories:
  - image-segmentation
language:
  - en
pretty_name: CSeg
size_categories:
  - 100K<n<1M

Dataset Card for Dataset Name

CSeg (Change Segmentation) dataset is a large-scale, varying-view, and text-guided change detection dataset. With over 500k datapoints, CSeg was used to train ViewDelta to associate a text prompt to relevant changes between two images.

Paper and Repo

• Repository: Github
• Paper: ViewDelta: Text-Prompted Change Detection in Unaligned Images

Dataset Download

wget https://huggingface.co/datasets/hoskerelab/CSeg/resolve/main/download.sh
bash download.sh

Dataset Structure

cseg_test.csv and cseg_train.csv have been provided to provide the paths of the corresponding Image A, Image B, and their Label.

cseg/
├── test/
│   ├── image_a
│   │   ├── 223762_1.png
│   │   ├── 223797_1.png
│   │   └── ...
│   ├── image_b
│   │   ├── 223762_1.png
│   │   ├── 223797_1.png
│   │   └── ...
│   └── label
│       ├── 223762_1.png
│       ├── 223797_1.png
│       └── ...
├── train/
│   ├── image_a
│   │   ├── 223750_1.png
│   │   ├── 223751_1.png
│   │   └── ...
│   ├── image_b
│   │   ├── 223750_1.png
│   │   ├── 223751_1.png
│   │   └── ...
│   └── label
│       ├── 223750_1.png
│       ├── 223751_1.png
│       └── ...
├── cseg_test.csv
└── cseg_train.csv

Citation

@misc{varghese2025viewdeltatextpromptedchangedetection,
      title={ViewDelta: Text-Prompted Change Detection in Unaligned Images}, 
      author={Subin Varghese and Joshua Gao and Vedhus Hoskere},
      year={2025},
      eprint={2412.07612},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.07612}, 
}