Peanuttoad/StreamGaze_v2

StreamGaze是一个全面的流媒体视频基准，用于评估多模态大型语言模型（MLLMs）在基于注视的问答任务上的表现，涵盖过去、现在和未来上下文。数据集包含多种注视相关的任务类别，如注视序列匹配、非注视对象识别、对象转换预测、场景回忆、对象识别（简单/困难）、对象属性识别、未来动作预测、注视触发警报和对象出现警报。数据集结构包括元数据、问答文件和视频文件。元数据CSV文件包含视频标识符、注视段ID、时间边界、注视中心坐标、主要对象等信息。问答JSON文件包含问题、时间戳、答案和选项等信息。

StreamGaze is a comprehensive streaming video benchmark for evaluating MLLMs on gaze-based QA tasks across past, present, and future contexts. The dataset includes various gaze-related task categories such as Gaze Sequence Matching, Non-Fixated Object Identification, Object Transition Prediction, Scene Recall, Object Identification (Easy/Hard), Object Attribute Recognition, Future Action Prediction, Gaze-Triggered Alert, and Object Appearance Alert. The dataset structure includes metadata, QA files, and video files. Metadata CSV files contain video identifiers, fixation segment IDs, temporal boundaries, gaze center coordinates, primary objects, etc. QA JSON files include questions, timestamps, answers, and options.