CDLA / README.md
SWHL's picture
Update README.md
f4b1146 verified
metadata
license: apache-2.0
language:
  - zh
size_categories:
  - 1K<n<10K

CDLA: A Chinese document layout analysis (CDLA) dataset

介绍

CDLA是一个中文文档版面分析数据集,面向中文文献类(论文)场景。包含以下10个label:

正文 标题 图片 图片标题 表格 表格标题 页眉 页脚 注释 公式
Text Title Figure Figure caption Table Table caption Header Footer Reference Equation

共包含5000张训练集和1000张验证集,分别在train和val目录下。

整理自:CDLA

标注可视化:

使用方式

from datasets import load_dataset

dataset = load_dataset("SWHL/CDLA")

train_data = dataset["train"]
print(train_data[0])

val_data = dataset["validation"]
print(val_data[0])

# {'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=1240x1754 at 0x12FEE3DF0>,
# 'version': '4.5.6', 'flags': {},
# 'shapes': [
#     {'label': 'Header', 'points': [[118.0, 135.66666666666669]], 'group_id': None, 'shape_type': 'polygon', 'flags': {}}
# ],
# 'imagePath': 'train_0001.jpg', 'imageData': None, 'imageHeight': 1754, 'imageWidth': 1240}

下载链接

  • 百度云下载:link, 提取码: tp0d
  • Google Drive Download:link

标注格式

我们的标注工具是labelme,所以标注格式和labelme格式一致。这里说明一下比较重要的字段:

  • shapes: shapes字段是一个list,里面有多个dict,每个dict代表一个标注实例。
  • labels: 类别。
  • points: 实例标注。因为我们的标注是Polygon形式,所以points里的坐标数量可能大于4。
  • shape_type: "polygon"
  • imagePath: 图片路径/名
  • imageHeight: 高
  • imageWidth: 宽

展示一个完整的标注样例:

{
  "version":"4.5.6",
  "flags":{},
  "shapes":[
    {
      "label":"Title",
      "points":[
        [
          553.1111111111111,
          166.59259259259258
        ],
        [
          553.1111111111111,
          198.59259259259258
        ],
        [
          686.1111111111111,
          198.59259259259258
        ],
        [
          686.1111111111111,
          166.59259259259258
        ]
      ],
      "group_id":null,
      "shape_type":"polygon",
      "flags":{}
    },
    {
      "label":"Text",
      "points":[
        [
          250.5925925925925,
          298.0740740740741
        ],
        [
          250.5925925925925,
          345.0740740740741
        ],
        [
          188.5925925925925,
          345.0740740740741
        ],
        [
          188.5925925925925,
          410.0740740740741
        ],
        [
          188.5925925925925,
          456.0740740740741
        ],
        [
          324.5925925925925,
          456.0740740740741
        ],
        [
          324.5925925925925,
          410.0740740740741
        ],
        [
          1051.5925925925926,
          410.0740740740741
        ],
        [
          1051.5925925925926,
          345.0740740740741
        ],
        [
          1052.5925925925926,
          345.0740740740741
        ],
        [
          1052.5925925925926,
          298.0740740740741
        ]
      ],
      "group_id":null,
      "shape_type":"polygon",
      "flags":{}
    },
    {
      "label":"Footer",
      "points":[
        [
          1033.7407407407406,
          1634.5185185185185
        ],
        [
          1033.7407407407406,
          1646.5185185185185
        ],
        [
          1052.7407407407406,
          1646.5185185185185
        ],
        [
          1052.7407407407406,
          1634.5185185185185
        ]
      ],
      "group_id":null,
      "shape_type":"polygon",
      "flags":{}
    }
  ],
  "imagePath":"val_0031.jpg",
  "imageData":null,
  "imageHeight":1754,
  "imageWidth":1240
}

转COCO格式

# train
python3 labelme2coco.py CDLA_dir/train train_save_path  --labels labels.txt

# val
python3 labelme2coco.py CDLA_dir/val val_save_path  --labels labels.txt

转换结果保存在train_save_path/val_save_path目录下。

labelme2coco.py取自labelme,更多信息请参考labelme官方项目