SAV-Caption
收藏资源简介:
SAV-Caption 数据集是 VoCap 模型的训练数据集,由 SAV 数据集的分割掩码和自动生成的对象描述组成。SAV-Caption 训练集包含约 50,000 个视频,每个视频都有一个与之对应的对象描述。SAV-Caption 验证集包含 155 个视频,每个视频的每个对象都有三个不同的手动描述。该数据集旨在解决视频对象分割和描述问题,并为视频对象分割和描述任务提供了新的基准。
The SAV-Caption dataset, the training dataset for the VoCap model, consists of segmentation masks derived from the SAV dataset and automatically generated object descriptions. The SAV-Caption training set contains approximately 50,000 videos, each paired with a corresponding object description. The SAV-Caption validation set includes 155 videos, with three distinct manual descriptions for each object in every video. This dataset is designed to tackle the tasks of video object segmentation and description, and establishes a novel benchmark for these tasks.
数据集概述
基本信息
- 数据集名称: vocap
- 来源论文: VoCap: Video Object Captioning and Segmentation from Any Prompt
- 论文链接: https://arxiv.org/abs/2508.21809
- 数据内容: 包含SAV数据集中对象的伪标注和人工标注描述
数据组成
验证集标注
- 文件: sav_caption_val_human.csv
- 标注方式: 人工标注
- 标注详情: 每个标注对象由三位不同标注人员分别提供描述
- 数据特点: 大多数video_id与object_id组合重复出现三次
训练集标注
- 文件: sav_caption_train_automatic.csv(14MB)
- 标注方式: 自动生成(使用Gemini 1.5 Pro模型)
- 生成方法: 基于真实标注突出显示感兴趣对象后生成以对象为中心的描述
数据格式
- 文件类型: CSV文件(带表头)
- 字段组成: video_id, object_id, caption(逗号分隔)
- 数据记录: 每行包含视频ID、对象ID和对应的描述文本
引用信息
bibtex @inproceedings{uijings25vocap, title={{VoCap}: Video Object Captioning and Segmentation from Any Prompt}, author={Jasper Uijlings and Xingyi Zhou and Xiuye Gu and Arsha Nagrani and Anurag Arnab and Alireza Fathi and David Ross and Cordelia Schmid}, booktitle={ArXiv}, year={2025}, }
许可信息
- 软件许可: Apache License 2.0
- 许可链接: https://www.apache.org/licenses/LICENSE-2.0
- 其他材料许可: Creative Commons Attribution 4.0 International License (CC-BY)
- 许可链接: https://creativecommons.org/licenses/by/4.0/legalcode
- 免责声明: 非Google官方产品,按"原样"提供,不提供任何明示或暗示的担保

- 1VoCap: Video Object Captioning and Segmentation from Any PromptGoogle DeepMind · 2025年



