DenseWorld-1M
收藏资源简介:
DenseWorld-1M是一个大规模、详细、密集的视觉语言数据集,用于多模态大型语言模型(MLLMs)的预训练、监督微调和后训练。该数据集包含实体级别的掩码和标签,对象级别的详细描述,以及场景级别的密集地描述。DenseWorld-1M数据集的创建过程分为三个阶段:开放世界感知、详细对象描述生成和密集描述合并。DenseWorld-1M数据集旨在解决现有数据集缺乏详细、密集的对象描述、地描述和对象位置信息的问题,并为下游任务如视觉语言理解、视觉定位和区域描述生成提供数据支持。
DenseWorld-1M is a large-scale, detailed, dense vision-language dataset tailored for pre-training, supervised fine-tuning, and post-training of multimodal large language models (MLLMs). This dataset encompasses entity-level masks and labels, granular object-level descriptions, and dense scene-level descriptions. The construction of the DenseWorld-1M dataset comprises three sequential stages: open-world perception, detailed object description generation, and dense description merging. The DenseWorld-1M dataset aims to mitigate the limitations of existing datasets, which lack detailed and dense object descriptions, scene-level descriptions, and precise object location information, while providing data support for downstream tasks including vision-language understanding, visual localization, and region description generation.
DenseWorld-1M 数据集概述
基本信息
- 数据集名称: DenseWorld-1M
- 研究领域: 计算机视觉、多模态大语言模型(MLLMs)
- 贡献机构: Bytedance Seed、武汉大学、北京大学
- 相关论文: arXiv链接
- 数据集状态: 清理中,预计2025年7月底前在HuggingFace开源
数据集特点
- 规模: 大规模(百万级)
- 标注类型: 密集、详细、基于实体的描述
- 标注内容:
- 实体级掩码和标签
- 对象级详细描述
- 空间和关系密集描述
标注流程
- 开放世界感知: 获取实体级掩码和标签
- 详细对象描述生成: 基于掩码和标签生成对象级详细描述
- 密集描述合并: 将对象描述和掩码合并为空间和关系密集描述
辅助工具
- 视觉语言模型(VLM):
- Detailed Region Caption模型
- Spatial Caption Merging模型
应用场景
- 视觉语言理解
- 视觉定位
- 区域描述生成
引用信息
bibtex @misc{li2025denseworld1m, title={DenseWorld-1M: Towards Detailed Dense Grounded Caption in the Real World}, author={Xiangtai Li and Tao Zhang and Yanwei Li and Haobo Yuan and Shihao Chen and Yikang Zhou and Jiahao Meng and Yueyi Sun and Shilin Xu and Lu Qi and Tianheng Cheng and Yi Lin and Zilong Huang and Wenhao Huang and Jiashi Feng and Guang Shi}, year={2025}, eprint={2506.24102}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2506.24102}, }




