InterVLA
收藏资源简介:
InterVLA是一个大规模的人-物-人交互数据集,涵盖了多样化的交互模式和稳定的自我视角感知。该数据集包含3.9K序列,总时长11.4小时,共1.2M帧的多模态交互数据,包括自我视角和外部视角的RGB视频、语言命令和高精度的人/物运动。InterVLA数据集的收集过程是在视觉-语言-行动框架下进行的,其中助手根据自我视角的视觉和语言命令为导师提供服务。数据集的创建过程包括场景设置、数据收集、数据处理等多个环节,旨在促进通用人工智能的发展。
InterVLA is a large-scale human-object-human interaction dataset that encompasses diverse interaction patterns and robust egocentric perception. This dataset comprises 3.9K sequences with a total duration of 11.4 hours and 1.2 million frames of multimodal interaction data, including RGB videos from both egocentric and exocentric perspectives, linguistic commands, and high-precision human and object motion data. The collection of the InterVLA dataset is carried out under the vision-language-action framework, where an assistant serves a tutor based on egocentric visual inputs and linguistic commands. The creation of this dataset involves multiple stages including scene setup, data collection, and data processing, with the goal of advancing the development of artificial general intelligence (AGI).



