Ref-SAV
收藏资源简介:
Ref-SAV数据集是一个用于视频对象分割的自动标注数据集,包含超过72,000条对象表达,涵盖了复杂视频场景中的多种对象。该数据集由字节跳动种子团队创建,旨在提升模型在复杂环境中的视频对象分割性能。数据集通过自动标注流程生成,并手动验证了2,000个视频对象,以确保数据质量。Ref-SAV数据集的应用领域主要集中在视频理解、对象分割和视觉问答等任务,旨在解决复杂场景下的视频对象分割问题。
Ref-SAV is an automatically annotated dataset for video object segmentation, which contains over 72,000 object expressions covering various objects in complex video scenarios. Developed by the ByteDance Seed Team, this dataset is designed to enhance the performance of models in video object segmentation tasks within complex environments. It is generated via an automatic annotation workflow, with 2,000 video objects manually verified to guarantee data quality. The Ref-SAV dataset is primarily applied to tasks such as video understanding, object segmentation, and visual question answering, aiming to solve the challenges of video object segmentation in complex scenes.

- 1Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos加州大学默塞德分校, 字节跳动种子, 武汉大学, 北京大学 · 2025年



