Long-Horizon Office Activity Dataset
收藏资源简介:
该数据集由卡内基梅隆大学等机构创建,旨在为长时程人类活动模拟的行为逼真度评估提供基准。数据集包含43小时多摄像头视频,记录了一间共享办公室中55名人员的自然活动,其中5名核心人员被连续观测数天,每人平均29小时。数据通过11个同步摄像头连续录制一周,经人工标注转化为包含6种活动标签(如使用电脑、交谈、用餐等)的轨迹序列。该数据集用于衡量LLM模拟器在活动级、时段级和日级时间粒度以及个体与群体分析层次上的行为逼真度,揭示了不同条件方法在还原真实行为模式上的差异与局限。
This dataset was developed by Carnegie Mellon University and other institutions to provide a benchmark for behavioral fidelity evaluation in long-duration human activity simulation. It contains 43 hours of multi-camera video footage documenting natural activities of 55 individuals in a shared office, among which 5 core participants were monitored continuously for several days, with an average of 29 hours per person. Recorded over one week using 11 synchronized cameras, the data was manually annotated into trajectory sequences with 6 activity labels, such as using computers, having conversations, dining, and more. This dataset is utilized to assess the behavioral fidelity of LLM simulators across three temporal granularities (activity-level, session-level, and day-level) and two analytical levels (individual and group), uncovering the differences and limitations of different conditional approaches in reproducing real-world behavioral patterns.

- 1Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations卡内基梅隆大学·人机交互研究所; 富士通株式会社; 富士通美国研究院 · 2026年



