velvet-pine-22/PULSE

PULSE是一个同步五模态数据集，用于多模态日常活动理解。它包含来自40名志愿者在9种场景下录制的数据，涵盖五种硬件同步在100 Hz的模态：运动捕捉（MoCap）、肌电图（EMG）、眼动追踪（EyeTrack）、惯性测量单元（IMU）和指尖压力（Pressure）。数据集包含337条录制（304条任务录制和33条运动原语录制），总时长约9.7小时，并提供了7,789个密集标注的动作片段，每个片段标注了运动原语（17种观察到的类型）、手部信息（左/右/双手）、操作对象（57种独特对象）和自然语言描述。数据集支持多种基准任务，包括场景识别、细粒度动作识别、抓取起始预测、缺失模态鲁棒性、触觉驱动的抓取状态识别和跨模态压力预测。数据以CSV和JSON格式提供，并包含训练/测试划分（基于志愿者独立划分）。数据集遵循CC BY-NC 4.0许可证，并附有使用限制，禁止商业重新部署、参与者重新识别以及未经同意的监控或生物识别应用。PULSE is a synchronized five-modal dataset for multimodal daily activity understanding. It contains data recorded from 40 volunteers across 9 scenarios, covering five modalities synchronized at 100 Hz: motion capture (MoCap), electromyography (EMG), eye tracking (EyeTrack), inertial measurement unit (IMU), and fingertip pressure (Pressure). The dataset includes 337 recordings (304 task recordings and 33 motion primitive recordings), with a total duration of approximately 9.7 hours, and provides 7,789 densely annotated action segments. Each segment is annotated with motion primitives (17 observed types), hand information (left/right/both hands), manipulated objects (57 unique objects), and natural language descriptions. The dataset supports a variety of benchmark tasks, including scene recognition, fine-grained action recognition, grasp initiation prediction, missing modality robustness, tactile-driven grasp state recognition, and cross-modal pressure prediction. The data is provided in CSV and JSON formats, and includes training/test splits based on independent volunteer partitioning. The dataset follows the CC BY-NC 4.0 license, with usage restrictions prohibiting commercial redistribution, participant re-identification, and unauthorized surveillance or biometric applications.