Spatial-mem
收藏资源简介:
Spatial-mem数据集是为训练和评估一种新型视频世界模型框架而创建的,该框架旨在通过基于几何的长久空间记忆来增强视频世界模型的长久一致性。该数据集包含静态场景部分的点云表示,用于存储和检索信息,并通过过滤动态元素来记住静态场景。数据集还包含一组历史参考帧作为稀疏的长久情节记忆。该数据集旨在帮助模型在长时间范围内保持场景一致性,从而为计算机图形学、机器人学和其他交互式应用提供无限长度且一致的世界生成能力。
The Spatial-mem dataset was created for training and evaluating a novel video world model framework, which aims to enhance the long-term consistency of video world models through geometry-based persistent spatial memory. This dataset includes point cloud representations of static scene segments, used for storing and retrieving information, and remembers static scenes by filtering out dynamic elements. It also contains a set of historical reference frames as sparse persistent episodic memory. This dataset is designed to help models maintain scene consistency over long time horizons, thereby providing infinite-length and consistent world generation capabilities for computer graphics, robotics, and other interactive applications.




