TVR-Ranking
收藏资源简介:
TVR-Ranking数据集由南洋理工大学的研究团队开发,基于TVR数据集的视频片段和时刻注释,包含94,442个查询-时刻对。数据集通过人工注释相关性级别,支持排名视频时刻检索任务。数据集的创建过程中,研究者通过ChatGPT生成的提示词替换原始时刻描述中的角色名称,以生成不精确的查询。该数据集主要应用于视频时刻检索领域,旨在通过不精确的查询从视频集合中检索并排名相关时刻,解决实际搜索场景中的问题。
The TVR-Ranking dataset was developed by the research team from Nanyang Technological University. Built upon the video clips and temporal annotations of the original TVR dataset, it contains 94,442 query-moment pairs. The dataset supports the ranked video moment retrieval task with manually annotated relevance levels. During its construction, researchers replaced character names in the original moment descriptions with prompts generated by ChatGPT to generate imprecise queries. This dataset is primarily utilized in the field of video moment retrieval, aiming to retrieve and rank relevant moments from video collections via imprecise queries, thereby solving practical problems in real-world search scenarios.
Video Moment Retrieval in Practical Setting: A Dataset of Ranked Moments for Imprecise Queries
数据集概述
该数据集是为论文《Video Moment Retrieval in Practical Settings: A Dataset of Ranked Moments for Imprecise Queries》创建的基准数据集。
数据集结构
数据集包含以下文件:
val.jsontest.jsontrain_top01.jsontrain_top20.jsontrain_top40.jsonvideo_corpus.json
特征文件
数据集包括以下特征文件:
- 查询BERT特征
- 视频和字幕特征
实验结果
基线模型性能
| Model | Train Set Top N | IoU=0.3 | IoU=0.5 | IoU=0.7 |
|---|---|---|---|---|
| XML | 1 | 0.1010 | 0.0737 | 0.0258 |
| 20 | 0.2331 | 0.1700 | 0.0627 | |
| 40 | 0.2114 | 0.1530 | 0.0583 | |
| CONQUER | 1 | 0.0952 | 0.0808 | 0.0526 |
| 20 | 0.2130 | 0.1976 | 0.1527 | |
| 40 | 0.2183 | 0.2022 | 0.1524 | |
| ReLoCLNet | 1 | 0.1504 | 0.1303 | 0.0866 |
| 20 | 0.3815 | 0.3462 | 0.2381 | |
| 40 | 0.4418 | 0.4060 | 0.2787 |
ReLoCLNet性能
| Model | Train Set Top N | IoU=0.3 | IoU=0.5 | IoU=0.7 |
|---|---|---|---|---|
| NDCG@10 | ||||
| ReLoCLNet | 1 | 0.1575 | 0.1358 | 0.0908 |
| 20 | 0.3751 | 0.3407 | 0.2316 | |
| 40 | 0.4339 | 0.3984 | 0.2693 | |
| NDCG@20 | ||||
| ReLoCLNet | 1 | 0.1504 | 0.1303 | 0.0866 |
| 20 | 0.3815 | 0.3462 | 0.2381 | |
| 40 | 0.4418 | 0.4060 | 0.2787 | |
| NDCG@40 | ||||
| ReLoCLNet | 1 | 0.1533 | 0.1321 | 0.0878 |
| 20 | 0.4039 | 0.3656 | 0.2542 | |
| 40 | 0.4725 | 0.4337 | 0.3015 |
引用
如果该数据集对您的研究有帮助,请引用我们的工作:
@misc{liang2024tvrrankingdatasetrankedvideo, title={TVR-Ranking: A Dataset for Ranked Video Moment Retrieval with Imprecise Queries}, author={Renjie Liang and Li Li and Chongzhi Zhang and Jing Wang and Xizhou Zhu and Aixin Sun}, year={2024}, eprint={2407.06597}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2407.06597}, }

- 1TVR-Ranking: A Dataset for Ranked Video Moment Retrieval with Imprecise Queries南洋理工大学 · 2024年



