Tera-AI/STRIDE

STRIDE（时空道路图像探索数据集）是由大约820亿个标记组成的数据集，这些标记被排列成600万个视觉“句子”或标记序列。这些序列是从相对较小的131k全景图像集以及它们的元数据和公开可用的公路系统数据生成的。这种方法使信息量增加了27倍，允许进行生成世界模型的训练。本质上，STRIDE将真实的街景全景观测转化为一个可在空间和时间上自由游玩的可导航、交互式环境。

We develop STRIDE (Spatio-Temporal Road Image Dataset for Exploration), that consists of approximately 82B tokens which were arranged into a total of 6M visual sentences or token sequences. The sequences are generated from a relatively small set of 131k panoramic images, along with their metadata and openly available highway system data. Our approach enables a 27x increase in information, allowing for generative world model training. In essence, STRIDE turns real street-view panoramic observations into a navigable, interactive environment suitable for free-play across space and time.