Seed42Lab/RefOI

RefOI数据集旨在解决现有视觉-语言系统中指代表达式生成（REG）基准测试的局限性。该数据集基于OpenImages V7实例分割验证集构建，包含1485个真实世界对象实例，均匀分布在COCO和非COCO类别中。每个实例都标注了三个书面和两个口语的人称指代表达式，为评估视觉-语言模型的语用能力提供了丰富的资源。数据集分为“单一存在”和“共现”类别，允许进行细粒度的REG/REC性能分析。README还包含数据集模式、使用示例和实验结果的描述，比较了各种模型与人类生成的描述。

The RefOI dataset is designed to address limitations in existing benchmarks for Referring Expression Generation (REG) in vision-language systems. The dataset is built from the OpenImages V7 Instance Segmentation validation set and includes 1,485 real-world object instances, equally distributed across COCO and non-COCO classes. Each instance is annotated with three written and two spoken human referring expressions, providing a rich resource for evaluating the pragmatic competence of vision-language models. The dataset is split into single_presence and co-occurrence categories, allowing for fine-grained analysis of REG/REC performance. The README also includes a dataset schema, usage examples, and experimental results comparing various models against human-generated descriptions.