Paper2Fig100k

来自研究论文的超过100k个图形图像和文本标题的数据集。图形图像显示我n arXiv.org的研究论文的图表、方法和架构。我们还为每个图形提供文本标题，以及对图形的OCR检测和识别 (边界框和文本)。数据集结构由一个名为 “数字” 的目录和两个JSON文件 (训练和测试) 组成，其中包含每个图形的数据。每个JSON对象包含有关图形的以下信息: figure_id: 基于arXiv标识符的图形标识: <yymm>。<xxxxxx>-图形 <I>-<k>.png.ca选项: 从与图形相关的论文中提取的文本对。例如，该图的实际标题或对手稿中该图的引用。ocr_result: 在图像上执行OCR文本识别的结果。我们提供图像中存在的三胞胎 (边界框，置信度，文本) 的列表。方面: 图像的纵横比 (H/W)。A dataset containing over 100,000 graphical images and their corresponding text captions sourced from research papers. The graphical images depict figures, methods, and architectures from research papers hosted on arXiv.org. We also provide the text caption for each figure, along with OCR detection and recognition results for the figure (including bounding boxes and recognized text). The dataset structure consists of a directory named "digits" and two JSON files (for training and testing) that contain the data for each figure. Each JSON object contains the following information about the figure: figure_id: The figure identifier based on the arXiv identifier, formatted as <yymm>.<xxxxxx>-figure <I>-<k>.png. caption_pairs: Text pairs extracted from the paper associated with the figure, such as the actual caption of the figure or a reference to the figure within the manuscript. ocr_result: The results of OCR text recognition performed on the image. We provide a list of triplets (bounding box, confidence score, recognized text) present in the image. aspect_ratio: The aspect ratio of the image (H/W).

搜集汇总

背景与挑战

背景概述

Paper2Fig100k数据集包含超过10万个来自arXiv研究论文的图形图像及其文本标题，涵盖图表、方法和架构等内容，并提供OCR检测结果。该数据集由计算机视觉中心等机构于2022年11月发布，结构包括图像目录和JSON文件，用于训练和测试任务。