ScribbleSimulation

This repository is created for reimplementation of the ScribbleCOCO and ScribbleCityscapes. The codes are available at BaiduNetdisk. The extraction code is in the paper.

Terms of use

• ScribbleCOCO: We follow the same license as COCO dataset for images.
• ScribbleCityscapes: We follow the same license as The Cityscapes Dataset.
• The datasets cannot be used for commercial purposes. The datasets are created for research purposes.

Environment

• python>=3.9
• opencv-python
• matplotlib
• skimage
• multiprocessing

Dataset preparation

Download the COCO dataset from the official webset.

dataset/COCO2014
├── annotations
├── test2014
├── train2014
└── val2014

Example generating ScribbleCOCO

Run with the script: generate_coco.sh to generate scribble masks 3 times with randomness.

The allowed args are :

parser.add_argument('--dataDir',default='./COCO2014',type=str)
parser.add_argument('--dataType',default='train2014',type=str)
parser.add_argument('--save_dir',default='./coco2014_train_scribble_random',type=str)
parser.add_argument('--random',default='True',type=str,help='The generate the scribble with a random walk path, or chose the longest path as the scribble')
parser.add_argument('--numworkers',default=1,type=int)

ScribbleCOCO

The complete dataset is avaliable at BaiduNetdisk. The ScribbleCOCO need at leat 24G space. The dataset is recommended to untar with the following structure:

ScribbleCOCO/
├── ImageSets/SegmentationAug
    ├──train.txt
    └──val.txt
├── coco2014_train_scribble_r1(The scribble masks, png files, about 923M.)
├── coco2014_train_scribble_r2
├── coco2014_train_scribble_r3
├── JPEGImages (The jpeg image files, about 19G)
└── SegmentationClassAug  (The ground truth masks, png files, about 1.1G)
(The followings are Optional)
├── pseudolabels (4.8G in total.)
    ├── toco (Pseudo label masks frome the original ToCo, png files, about 429M)
    ├── toco_r1 (Pseudo label masks from our scribble-promoted ToCo, png files, about 441M)
    ├── toco_r2
    ├── toco_r3
    ├── cutmix (Pseudo labels masks from the Cutmix, png files, about 418M)
    ├── has (Pseudo labels masks from the Has, png files, about 415m)
    └── recam (Pseudo labels masks from the ReCam, png files, about 360m)
├── scribble_dsmp (Distance maps of the scribbles. About 176G, recommend to generate from the code.)
└── pseudolabel_dsmp (Distance maps of the pseudo labels. About 160G, recommend to generate from the code.)

.

ScribbleCityscapes

A COCO-style Cityscapes is avaliable at BaiduNetdisk.(preserved for furture open-sourced.) Download the COCO-style Cityscapes ↑ The complete dataset of ScribbleCityscapes is available at BaiduNetdisk.

ScribbleCityscapes/
├── ImageSets/SegmentationAug
    ├──train.txt
    └──val.txt
├── cityscapes_scribble_r1(The scribble masks, png files, about 74M.)
├── cityscapes_scribble_r2
├── cityscapes_scribble_r3
├── JPEGImages (The jpeg image files, about 1.6G)
└── SegmentationClassAug  (The ground truth masks, png files, about 171M)
(The followings are Optional)
├── pseudolabels (76M in total.)
    ├── ToCoR1 (Pseudo label masks from our scribble-promoted ToCo, png files, about 26M)
    ├── ToCoR2
    └── ToCoR3
├── scribble_dsmp (Distance maps of the scribbles. About 13G, recommend to generate from the code.)
└── pseudolabel_dsmp (Distance maps of the pseudo labels. About 16G, recommend to generate from the code.)

ScribbleSup and ScribbleACDC

These two datasets are public datasets. ScribbleSup: The original scribble annotations were recorded as a serises of points, where you can find them in ../scribble_annotation/pascal_2012/*.xml. I recollected the ScribbleSup data in 2023 for scribble-supervised semantic segmentation. I convert them into the png files following this code by matlab. The complete dataset is availiable at google drive.

ScribbleACDC: The offical ACDC dataset is a 3D medical image segmentation for Automated Cardiac Diagnosis Challenge. Valvano et. al provided the human annotated scribble annotations in 2021. I here recollected this data following PASCAL VOC format. The complete dataset is availiable at BaiduNetdisk. The code is 4wtu.

Citation

If you found this repo is helpful, please cite:

@article{zhang2025exploiting,
  title={Exploiting Inherent Class Label: Towards Robust Scribble Supervised Semantic Segmentation},
  author={Zhang, Xinliang and Zhu, Lei and Zeng, Shuang and He, Hangzhou and Fu, Ourui and Yao, Zhengjian and Xie, Zhaoheng and Lu, Yanye},
  journal={arXiv preprint arXiv:2503.13895},
  year={2025}
}
@inproceedings{zhang2024scribble,
  title={Scribble hides class: Promoting scribble-based weakly-supervised semantic segmentation with its class label},
  author={Zhang, Xinliang and Zhu, Lei and He, Hangzhou and Jin, Lujia and Lu, Yanye},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={7},
  pages={7332--7340},
  year={2024}
}