metadata

license: cc-by-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_name
      dtype: string
    - name: width
      dtype: int64
    - name: height
      dtype: int64
    - name: instances
      list:
        - name: category_id
          dtype: int64
        - name: mask
          sequence:
            sequence: float64
  splits:
    - name: train
      num_bytes: 8927542
      num_examples: 200
    - name: validation
      num_bytes: 4722935
      num_examples: 100
    - name: test
      num_bytes: 3984722
      num_examples: 100
  download_size: 16709320
  dataset_size: 17635199

Line Graphics (LG) dataset

This is the official page for the LG dataset, as featured in our paper Line Graphics Digitization: A Step Towards Full Automation.

By Omar Moured et al.

Dataset Summary

The dataset includes instance segmentation masks for 400 real line chart images, manually labeled into 11 categories by professionals. These images were collected from 5 different professions to enhance diversity. In our paper, we studied two levels of segmentation: coarse-level, where we segmented (spines, axis-labels, legend, lines, titles), and fine-level, where we further segmented each category into x and y subclasses (except for legend and lines), and individually segmented each line.

Category ID Reference

class_id_mapping = {
    "Label": 0,
    "Legend": 1,
    "Line": 2,
    "Spine": 3,
    "Title": 4,
    "ptitle": 5,
    "xlabel": 6,
    "xspine": 7,
    "xtitle": 8,
    "ylabel": 9,
    "yspine": 10,
    "ytitle": 11
}

Dataset structure (train, validation, test)

image - contains the PIL image of the chart
image_name - image name with PNG extension
width - original image width
height - original image height
instances - contains n number of labeled instances, each instance dictionary has {category_id, annotations}. The annotations are in COCO format.

Sample Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("omoured/line-graphics-dataset")

# Access the training split
train_dataset = dataset["train"]

# Print sample data
print(dataset["train"][0])

You can render the masks using pycocotools library as follows:

from pycocotools import mask

polygon_coords = dataset['train'][0]['instances'][1]['mask']
image_width = dataset['validation'][0]['width']
image_height = dataset['validation'][0]['height']

mask_binary = mask.frPyObjects(polygon_coords, image_height, image_width) 

segmentation_mask = mask.decode(mask_binary)

Copyrights

This dataset is published under the CC-BY 4.0 license, which allows for unrestricted usage, but it should be cited when used.

Citation

@inproceedings{moured2023line,
  title={Line Graphics Digitization: A Step Towards Full Automation},
  author={Moured, Omar and Zhang, Jiaming and Roitberg, Alina and Schwarz, Thorsten and Stiefelhagen, Rainer},
  booktitle={International Conference on Document Analysis and Recognition},
  pages={438--453},
  year={2023},
  organization={Springer}
}

Contact

If you have any questions or need further assistance with this dataset, please feel free to contact us:

Omar Moured, [email protected]