ArT场景文本数据集 包括10166幅图像

ArT数据集将包括10166幅图像。它被分为一个包含5603张图像的训练集和一个包含4563张图片的测试集。 ArT是Total Text[4]、SCUT-CTW1500[5]和百度曲线场景文本的组合，收集这些文本的目的是将任意形状的文本问题引入场景文本社区。在现有图像（3055张）的基础上，将超过7111张图像添加到两个数据集的混合中，这使得ArT成为当今规模更大的场景文本数据集之一。ArT数据集中总共有10166张图像。它被分成一个包含5603幅图像的训练集和一个包含4563幅新收集图像的测试集。ArT数据集是在考虑到文本形状多样性的情况下收集的，因此所有现有文本形状（即水平、多方向和弯曲）在数据集中都有大量存在，这使其成为一个独特的数据集，因为大多数现有数据集[1、2、3]仅由水平和多方向文本实例主导。 ArT数据集中的文本实例用（a）四边形边界框、8、10和12个顶点多边形边界框（更多详细信息请参见任务选项卡）和（b）转录进行注释。这两种注释都满足了这一挑战提出的（a）文本检测、（b）识别和（c）文本定位任务。 数据结构： Training Set Test Set

The ArT dataset comprises 10,166 images in total, split into a training set containing 5,603 images and a test set containing 4,563 images. ArT is a combination of Total Text[4], SCUT-CTW1500[5], and Baidu's Curved Scene Text, which was curated to introduce arbitrary-shaped text problems into the scene text community. Based on the existing 3,055 images, more than 7,111 additional images were added to the amalgamation of the two initial datasets, rendering ArT one of the larger-scale scene text datasets currently available. The ArT dataset contains 10,166 images in total, split into a training set with 5,603 images and a test set with 4,563 newly-collected images. The ArT dataset was curated with consideration of the diversity of text shapes, such that all existing text modalities (i.e., horizontal, multi-oriented, and curved) are abundantly represented in the dataset, making it a unique benchmark, as most existing datasets [1, 2, 3] are predominantly composed of only horizontal and multi-oriented text instances. Text instances in the ArT dataset are annotated using (a) quadrilateral bounding boxes, polygonal bounding boxes with 8, 10, and 12 vertices (for further details, please refer to the Task Tab), and (b) ground-truth transcriptions. These two annotation types support all three tasks proposed in this challenge: (a) text detection, (b) text recognition, and (c) text localization. Data Structure: Training Set, Test Set

搜集汇总

背景与挑战

背景概述

ArT场景文本数据集是一个包含10166幅图像、总计5.59G的大规模数据集，专门用于场景文本检测和识别任务。数据集分为5603张图像的训练集和4563张图像的测试集，结合了Total Text、SCUT-CTW1500和百度曲线场景文本，覆盖水平、多方向和弯曲等多种文本形状，注释包括边界框和转录，适用于文本检测、识别和定位。