3D物理对象数据集

3D物理对象数据集是由马克斯普朗克智能系统研究所创建，包含超过一百万张物理3D对象的图像，涉及七个变化因素，如对象颜色、形状、大小和位置。为精确控制所有变化因素，研究团队构建了一个实验平台，其中对象由机器人手臂移动。此外，还提供了两个模拟实验设置的数据集，首次系统地研究了不同解耦方法在真实数据上的表现，以及模拟数据如何用于构建真实世界的更好表示。该数据集的应用领域包括视觉概念学习、序列建模、基于好奇心的探索以及强化学习中的领域适应等，旨在解决人工智能领域的关键挑战。

The 3D Physical Object Dataset was created by the Max Planck Institute for Intelligent Systems. It comprises over one million images of physical 3D objects, involving seven variation factors such as object color, shape, size and position. To accurately control all these variation factors, the research team built an experimental platform where objects are manipulated by a robotic arm. In addition, two datasets with simulated experimental settings are provided, which for the first time systematically study the performance of different disentanglement methods on real-world data, as well as how simulated data can be used to construct better representations of the real world. The application fields of this dataset include visual concept learning, sequence modeling, curiosity-driven exploration, domain adaptation in reinforcement learning and other related areas, aiming to address key challenges in the field of artificial intelligence.