GBC10M

GBC10M提供了一种新的图像标注模式，结合了长文本描述、区域描述和场景图的优点。它通过连接区域描述来创建统一的描述，同时提供类似于场景图的结构化信息。数据集包含超过1000万张图像，并提供了图像URL和JSON lines格式的标注，平均每张图像包含12.24个顶点、21.81条边、17.67条描述和533.98个单词。标注信息包括边界框信息和合成文本描述，每个样本都以`GbcGraph`的形式存储，包含顶点和边的信息。GBC10M支持图像到文本和文本到图像的任务，并提供标准化数据操作，可以通过HuggingFace `datasets`库直接加载。数据集的标注是在CC BY-NC 4.0许可下发布的，原始图像URL文本样本由Google发布，个人图像受其自身版权保护。

GBC10M introduces a novel image annotation paradigm that combines the strengths of long-text descriptions, regional descriptions and scene graphs. It generates unified descriptions by concatenating regional descriptions, while also providing structured information analogous to scene graphs. The dataset contains over 10 million images, with image URLs and annotations in JSON Lines format provided. On average, each image contains 12.24 vertices, 21.81 edges, 17.67 descriptions and 533.98 words. The annotations include bounding box information and synthetic text descriptions, with each sample stored in the form of `GbcGraph`, which contains vertex and edge information. GBC10M supports image-to-text and text-to-image tasks, and offers standardized data operations that can be directly loaded via the Hugging Face `datasets` library. The dataset's annotations are released under the CC BY-NC 4.0 license. The original text samples paired with image URLs are published by Google, and individual images are protected by their respective copyrights.