Open Images V4

Open Images V4 在多个维度上提供了大规模：3010 万个图像级标签用于 19800 个概念，1540 万个边界框用于 600 个对象类别，以及涉及 57 个类别的 375000 个视觉关系注释。特别是对于对象检测，提供的边界框比下一个最大的数据集（1.9M 图像上的 15.4M 框）多 15 倍。图像通常显示具有多个对象的复杂场景（平均每张图像 8 个带注释的对象）。它们之间的视觉关系被注释，这支持视觉关系检测，这是一项需要结构化推理的新兴任务。

Open Images V4 provides large-scale annotations across multiple dimensions: 30.1 million image-level labels for 19,800 concepts, 15.4 million bounding boxes for 600 object categories, and 375,000 visual relationship annotations covering 57 categories. Specifically for object detection, the provided bounding boxes are 15 times more than those of the next largest dataset, which has 15.4 million boxes across 1.9 million images. The images typically depict complex scenes with multiple objects, with an average of 8 annotated objects per image. Visual relationships between these objects are annotated, which supports visual relationship detection, an emerging task that requires structured reasoning.

搜集汇总

背景与挑战

背景概述

Open Images V4是一个大规模图像数据集，包含3010万个图像级标签、1540万个边界框和375000个视觉关系注释，支持图像分类、物体检测和视觉关系检测等任务。该数据集由Google Research于2020年发布，图像通常呈现复杂场景，平均每张图像有8个带注释的对象。