DIS5K (Dichotomous Image Segmentation 5K)

High-accuracy dichotomous image segmentation dataset, mirrored to the nobg org for convenience. Each example is an RGB image paired with a high-quality binary segmentation label.

• image: RGB source image
• label: binary segmentation ground-truth mask
• image_name: original image file name
• label_name: original label file name

Splits: DIS_TR (train, 3000), DIS_VD (validation, 470), and the four test sets DIS_TE1–DIS_TE4 (500 each, increasing in structural complexity).

Source & credit

Ported from DIS (github.com/xuebinqin/DIS), specifically from this drive link.

• 🔗 Dataset page: https://xuebinqin.github.io/dis/index.html
• 🔗 GitHub repo: https://github.com/xuebinqin/DIS

Citation

@InProceedings{qin2022,
  author={Xuebin Qin and Hang Dai and Xiaobin Hu and Deng-Ping Fan and Ling Shao and Luc Van Gool},
  title={Highly Accurate Dichotomous Image Segmentation},
  booktitle={ECCV},
  year={2022}
}