VistaDPO-7k
收藏资源简介:
VistaDPO-7k是一个大规模的空间-时间标注视频数据集,由新加坡国立大学构建,包含7.2K个问答对,每个问答对都包含选定的和拒绝的回答,以及时间戳、关键帧和边界框等空间-时间定位信息。该数据集旨在通过细粒度的视频-语言偏好对齐,减少视频语言失配和幻觉问题,支持多层次偏好优化,包括实例级、时间级和感知级,以增强大型视频模型的空间-时间推理能力。
VistaDPO-7k is a large-scale spatio-temporal annotated video dataset constructed by the National University of Singapore, which contains 7.2K question-answer pairs. Each pair includes both selected and rejected responses, as well as spatio-temporal localization information such as timestamps, keyframes and bounding boxes. This dataset aims to mitigate video-language misalignment and hallucination issues via fine-grained video-language preference alignment, supports multi-level preference optimization covering instance-level, temporal-level and perceptual-level, so as to enhance the spatio-temporal reasoning capabilities of large video models.

- 1VistaDPO: Video Hierarchical Spatial-Temporal Direct Preference Optimization for Large Video Models香港大学, 香港科技大学, 新加坡国立大学, 德克萨斯大学达拉斯分校, 南洋理工大学 · 2025年



