CHang/ROVI

ROVI是一个高质量的合成数据集，包含100万张经过精心挑选的网页图片，具有全面的图像描述和边界框注释。该数据集采用创新的VLM-LLM重描述策略，提供了开放词汇实例定位的文本到图像生成能力。数据集以JSON格式提供，包括训练和验证分割，每个样本包括图像URL、源信息、尺寸、边界框数据等。数据集还包括详细的管道概述，解释了VLM-LLM重描述策略的阶段以及所使用的开源模型。此外，还讨论了数据集的局限性、致谢信息、引用信息和联系方式。

ROVI is a high-quality synthetic dataset featuring 1M curated web images with comprehensive image descriptions and bounding box annotations. Using a novel VLM-LLM re-captioning strategy, ROVI enables open-vocabulary instance-grounded text-to-image generation. The dataset is structured in JSON format with train and validation splits, and each sample includes image URLs, source information, dimensions, bounding box data, and more. The dataset also includes a detailed pipeline overview, explaining the stages of the VLM-LLM re-captioning strategy and the open-source models used.