4D-controlled synthetic dataset
收藏资源简介:
该数据集由斯坦福大学和苏黎世联邦理工学院联合构建,旨在支持视频生成模型中时空解耦控制的研究。数据集包含通过Blender的PointOdyssey框架生成的合成动态场景,其中场景动态(世界时间)和摄像机轨迹被独立参数化,例如通过慢动作、暂停或随机速度变化生成时间重映射序列,并搭配多样化摄像机视角渲染。数据标注了精确的世界时间标签和摄像机参数,为4D可控视频扩散模型的训练提供了显式监督。其核心应用领域包括电影特效(如子弹时间)、游戏及XR场景中动态世界的自由视角探索。
This dataset was jointly constructed by Stanford University and ETH Zurich to support research on spatio-temporal decoupled control in video generation models. It comprises synthetic dynamic scenes generated via Blender's PointOdyssey framework, where scene dynamics (world time) and camera trajectories are independently parameterized. For example, time-remapped sequences are generated through slow motion, pausing, or random speed variations, paired with renderings from diverse camera viewpoints. The data is annotated with precise world time tags and camera parameters, providing explicit supervision for the training of 4D controllable video diffusion models. Its core application areas include film special effects (such as bullet time), free-viewpoint exploration of dynamic worlds in games and XR scenarios.




