Products for OCR and Information Extraction (POIE)
收藏资源简介:
POIE数据集是由华中科技大学和腾讯优图实验室合作创建的大型视觉信息提取数据集,包含3000张来自实际产品的营养成分标签的相机图像。该数据集不仅在布局、背景和字体上具有较大的变化,还包含了多达21种实体类型,其中一些实体有多种形式。POIE旨在解决真实世界中视觉信息提取的挑战,特别是在复杂布局和多变实体类型的情况下。数据集的创建过程涉及从产品图像中裁剪营养表,使用多个商业OCR引擎进行预标记,并通过人工校验和修复OCR错误。POIE的应用领域包括自动从视觉丰富的文档图像中提取结构化信息,如理解收据、商品和交通标志等。
The POIE dataset is a large-scale visual information extraction dataset jointly created by Huazhong University of Science and Technology and Tencent YouTu Lab. It contains 3000 camera-captured images of nutrition facts labels from real-world products. This dataset exhibits significant variations in layout, background, and font styles, and covers up to 21 entity types, some of which have multiple forms. POIE aims to address the challenges of visual information extraction in real-world scenarios, particularly those with complex layouts and diverse entity types. The dataset creation process involves cropping nutrition facts tables from product images, performing pre-labeling with multiple commercial OCR engines, and conducting manual verification and correction of OCR errors. Application fields of POIE include automated structured information extraction from visually-rich document images, such as interpreting receipts, commodities, traffic signs, and other similar contents.




