NTU RGB+D
收藏资源简介:
NTU RGB D是用于rgb-d人类动作识别的大规模数据集。它涉及从40名受试者中收集的60个动作类别的56,880个样本。动作一般可分为三类: 每日动作40项 (如喝酒、吃饭、读书),健康相关动作9项 (如打喷嚏、蹒跚、跌倒),相互动作11项 (如拳打脚踢、拥抱)。这些动作在对应于17个视频序列 (即S001-S017) 的17个不同场景条件下发生。使用三个具有不同水平成像视点的摄像机 (即-45,0和45) 捕获动作。提供了用于动作表征的多模态信息,包括深度图,3D骨架关节位置,RGB帧和红外序列。通过将40个受试者分为训练组和测试组的交叉受试者测试,以及使用一个摄像机 (45个) 进行测试和其他两个摄像机进行训练的交叉视图测试来执行性能评估。
NTU RGB-D is a large-scale dataset for RGB-D human action recognition. It contains 56,880 samples across 60 action categories collected from 40 subjects. Actions are generally categorized into three groups: 40 daily actions (e.g., drinking, eating, reading), 9 health-related actions (e.g., sneezing, staggering, falling), and 11 interactive actions (e.g., punching, kicking, hugging). These actions were captured under 17 distinct scene conditions corresponding to 17 video sequences (S001–S017). Three cameras with different horizontal viewing angles (-45°, 0° and 45°) were used to record the actions. Multimodal data for action representation are provided, including depth maps, 3D skeletal joint positions, RGB frames and infrared sequences. Performance evaluation is conducted using two standard protocols: cross-subject testing, where the 40 subjects are split into training and test subsets; and cross-view testing, where the camera with 45° viewing angle is used for testing while the other two cameras are employed for training.

- NTU RGB+D数据集首次发表,由台湾国立台湾大学发布,包含56880个视频序列,涵盖60个动作类别。
- NTU RGB+D数据集首次应用于动作识别研究,成为该领域的重要基准数据集之一。
- NTU RGB+D数据集的扩展版本NTU RGB+D 120发布,增加了60个新的动作类别,总视频序列数达到114480。
- NTU RGB+D数据集在多个国际会议和期刊上被广泛引用,推动了深度学习和计算机视觉领域的发展。
- NTU RGB+D数据集的跨视角版本NTU RGB+D X-Sub和NTU RGB+D X-View发布,进一步提升了数据集的应用价值。



