Nexdata/104320_Images_Korean_and_Hindi_OCR_Data_in_Natural_Scenes

---
license: cc-by-nc-nd-4.0
---
## Description
104,320 Images - Korean and Hindi OCR Data in Natural Scenes. The collecting scenes of this dataset include packaging, posters, tickets, reminders, menus, building signs, etc.. The data diversity includes multiple scenes, multiple shooting angles and multiple light conditions. For annotation, line-level polygon bounding box (or tetragon bounding box, rectangle bounding box) annotation, transcription and text attributes (language type) for the texts; vertical-level polygon bounding box (or tetragon bounding box, rectangle bounding box) annotation, transcription and text attributes (language type) for the text. The dataset can be used for Korean and Hindi OCR tasks in natural scenes.

For more details, please refer to the link: https://www.nexdata.ai/dataset/1254?source=Huggingface

## Data size
76,861 images of Korean, 555,913 bounding boxes; 27,459 images of Hindi, 200,453 bounding boxes
## Collecting environment
including packaging, posters, tickets, reminders, menus, building signs, etc.
## Data diversity
multiple natural scenes, multiple shooting angles, multiple light conditions
## Device
cellphone
## Collecting angle
looking up angle, looking down angle, eye-level angle
## Language distribution
Korean, Hindi, English (a few)
## Data format
the image data format is .jpg, the annotation file format is .json
## Bounding box shape distribution
315,822 tetragon bounding boxes and 240,091 polygon bounding boxes of Korean; 780 tetragon bounding boxes, 199,671 polygon bounding boxes and 2 rectangle bounding boxes of Hindi
## Annotation content
line-level polygon bounding box (or tetragon bounding box, rectangle bounding box) annotation, transcription and text attributes (language type) for the texts; vertical-level polygon bounding box (or tetragon bounding box, rectangle bounding box) annotation, transcription and text attributes (language type) for the text
## Accuracy
The error bound of each vertex of a bounding box is within 5 pixels, which is a qualified annotation, the accuracy of bounding boxes is not less than 95%; The texts transcription accuracy is not less than 95%.
# Licensing Information
Commercial License

原始信息汇总

数据集概述

数据集描述

• 图像数量：104,320张
• 语言：韩语和印地语
• 场景：包括包装、海报、票券、提醒、菜单、建筑标志等自然场景
• 数据多样性：多场景、多拍摄角度、多光线条件
• 标注类型：线级和垂直级多边形边界框（或四边形边界框、矩形边界框）标注，转录和文本属性（语言类型）

数据规模

• 韩语图像：76,861张，555,913个边界框
• 印地语图像：27,459张，200,453个边界框

收集环境

• 包括包装、海报、票券、提醒、菜单、建筑标志等

数据多样性

• 多自然场景、多拍摄角度、多光线条件

设备

• 手机

收集角度

• 仰视角度、俯视角度、水平角度

语言分布

• 韩语、印地语、英语（少量）

数据格式

• 图像格式：.jpg
• 标注文件格式：.json

边界框形状分布

• 韩语：315,822个四边形边界框，240,091个多边形边界框
• 印地语：780个四边形边界框，199,671个多边形边界框，2个矩形边界框

标注内容

• 线级和垂直级多边形边界框（或四边形边界框、矩形边界框）标注，转录和文本属性（语言类型）

准确性

• 边界框每个顶点的误差在5像素内，边界框准确率不低于95%
• 文本转录准确率不低于95%

许可信息

• 商业许可