ConfLab

ConfLab数据集是由代尔夫特理工大学的研究团队开发的一个新型概念，用于收集野外的自由站立社交对话的多模态多传感器数据。首次实施的ConfLab描述了在这里组织的一次国际会议上的专业网络活动，涉及48名会议参与者，数据集捕捉了状态、熟识度和网络动机的多样混合。捕获设置在保留隐私敏感性的同时改进了野外数据集的数据保真度：从非侵入性头顶视角拍摄的8个视频（1920×1080，60fps），以及具有板载记录身体运动（完整9轴IMU）、保护隐私的低频音频（1250Hz）和基于蓝牙的接近度的定制可穿戴传感器。此外，我们还开发了定制解决方案，用于在采集时进行分布式硬件同步，以及高效的时间连续注释身体关键点和动作的高采样率。我们的基准测试展示了一些与野外隐私保护社交数据分析相关的开放研究任务：从头顶摄像头视角检测关键点、基于骨骼的无音频说话者检测和F-formation检测。

The ConfLab Dataset is a novel resource developed by a research team at Delft University of Technology for collecting multimodal, multi-sensor data of freely-standing social conversations in the wild. The first instantiation of the ConfLab Dataset documents a professional networking event held at an international conference, involving 48 conference attendees, and captures a diverse mix of social states, familiarity levels, and networking motivations. The data capture setup enhances the data fidelity of in-the-wild datasets while preserving privacy sensitivity: eight videos shot from a non-intrusive overhead perspective (1920×1080, 60fps), alongside custom wearable sensors with on-board recording of body movements (full 9-axis IMU), privacy-preserving low-frequency audio (1250 Hz), and Bluetooth-based proximity tracking. Additionally, we developed custom solutions for distributed hardware synchronization during data collection, as well as high-sampling-rate, efficient temporally sequential annotation of body keypoints and actions. Our benchmark experiments showcase several open research tasks relevant to privacy-preserving social data analysis in the wild: keypoint detection from overhead camera views, skeleton-based audio-free speaker detection, and F-formation detection.