PubTables-1M: Detection Dataset

This is a Detection part of Microsoft PubTables-1M dataset, which is designed to address the limitations for table structure inference and extraction from unstructured documents. It comprises nearly one million tables extracted from scientific articles, offering support for multiple input modalities. Crucially, it includes detailed header and location information for table structures, enhancing its utility for diverse modeling approaches. The dataset not only quantifies improvements in training performance but also provides a more reliable estimate of model performance during evaluation for table structure recognition.

本数据集为微软PubTables-1M数据集的检测部分，旨在解决从非结构化文档中推断和提取表格结构的局限性。该数据集包含近百万张从科学文献中提取的表格，支持多种输入模式。尤为关键的是，它提供了详细的表头和结构位置信息，极大地提升了其在多样化建模方法中的应用价值。该数据集不仅量化了训练性能的提升，还为表格结构识别阶段的模型性能评估提供了更为可靠的估计。

搜集汇总

背景与挑战

背景概述

PubTables-1M: Detection Dataset 是一个用于表格检测的目标检测数据集，包含从科学文章中提取的近百万个表格，旨在解决非结构化文档中表格结构推断和提取的局限性。该数据集由575,305张图像组成，标注了683,056个对象，分为'table'和'table rotated'两个类别，并经过严格的注释验证和规范化处理以确保数据质量。数据集由Microsoft于2022年发布，提供训练、验证和测试分割，总大小约为76 GB，适用于表格提取研究。