Openstory++
收藏资源简介:
Openstory++是由华南理工大学、西湖大学等机构联合创建的大规模视觉故事叙述数据集,旨在通过实例级别的视觉分割注释增强故事连续性。该数据集包含超过1亿条高质量、完全注释的样本,特别强调在开放领域视频中提取关键帧,使用视觉-语言模型生成描述性字幕,并通过大型语言模型确保叙事连贯性。Openstory++不仅提供了丰富的开放领域资源,还通过自动化字幕生成和高分辨率图像,促进了多模态生成模型的发展,特别是在复杂叙事生成和开放领域环境中的应用。
Openstory++ is a large-scale visual storytelling dataset jointly created by South China University of Technology, Westlake University and other institutions, aiming to enhance story continuity through instance-level visual segmentation annotations. This dataset contains over 100 million high-quality, fully annotated samples, with a special focus on extracting key frames from open-domain videos, generating descriptive captions using vision-language models, and ensuring narrative coherence via large language models. Openstory++ not only provides abundant open-domain resources, but also promotes the development of multimodal generative models through automated caption generation and high-resolution images, especially for applications in complex narrative generation and open-domain scenarios.

- 1Openstory++: A Large-scale Dataset and Benchmark for Instance-aware Open-domain Visual Storytelling华南理工大学 西湖大学 中国科学院自动化研究所基础模型研究中心 阿卜杜拉国王科技大学 · 2024年



