cyberagent/OTR

OTR（文本覆盖移除）是一个合成的基准数据集，旨在推进图像中文本移除的研究。该数据集包含复杂的对象感知文本覆盖，具有干净无瑕疵的地面真实图像，使得评估场景超越了传统的场景文本数据集。数据集分为三个子集：OTR-easy测试集，OTR-hard测试集和训练集。OTR-easy包含在简单背景上渲染的文本的图像，使得文本区域的修复更加容易；OTR-hard包含在复杂结构对象上渲染的文本的图像，使得修复这些区域更加困难；训练集包含了来自两个源混合的图像。

OTR (Overlay Text Removal) is a synthetic benchmark dataset designed to advance research of text removal from images. It features complex, object-aware text overlays with clean, artifact-free ground truth images, enabling more challenging evaluation scenarios beyond traditional scene text datasets. The dataset consists of three subsets: the OTR-easy test set, the OTR-hard test set, and the training set. OTR-easy contains images with text rendered mostly on background regions with simple appearances, making text region inpainting easier; OTR-hard contains images with text rendered mostly over objects with complex structures, making it harder to naturally and seamlessly inpaint such regions; the training set contains a mix of images from both sources.