SoccerNet, HoloAssist
收藏资源简介:
本研究介绍了一个名为TGLG(时间相关语言生成)的新基准任务,用于评估视觉语言模型在实时交互环境中的感知更新和情境意识能力。为了支持这一基准,我们从体育广播和第一人称视角人机交互领域创建了数据集。SoccerNet数据集用于测试感知更新能力,包含16487条数据。HoloAssist数据集用于测试情境意识能力。这些数据集由时间戳标记的语句组成,以评估模型在实时视频流中生成语义准确且时间精确的语句的能力。
This study introduces a novel benchmark task named TGLG (Time-related Language Generation) for evaluating the perceptual updating and situational awareness capabilities of vision-language models in real-time interactive environments. To support this benchmark, we construct datasets from the domains of sports broadcasting and first-person human-computer interaction. The SoccerNet dataset, containing 16,487 samples, is utilized to test perceptual updating capabilities, while the HoloAssist dataset is designed to assess situational awareness. These datasets consist of timestamped utterances, aiming to evaluate a model's ability to generate semantically accurate and temporally precise statements in real-time video streams.




