Shadow-Sunlight, Shadow-Pointlight

本文介绍了两个新的数据集：Shadow-Sunlight和Shadow-Pointlight，由南加州大学信息科学研究所创建，旨在为因果表示学习提供更具挑战性的数据。这两个数据集模拟了不同光源类型（阳光和点光源）下的物体与阴影之间的因果关系，包含比现有数据集更多的生成因素，从而能够构建更复杂的因果图。数据集通过Blender软件生成，应用了Cycle渲染引擎，模拟了光、物体、地板和阴影之间的复杂交互。这些数据集适用于评估因果表示学习方法，特别是在处理计算机视觉问题如分布偏移、领域适应和公平性方面。

This paper introduces two novel datasets: Shadow-Sunlight and Shadow-Pointlight, developed by the Information Sciences Institute of the University of Southern California, which are designed to provide more challenging data for causal representation learning. These two datasets simulate the causal relationships between objects and their shadows under different light source types (sunlight and point light), and include more generative factors than existing datasets, enabling the construction of more complex causal graphs. The datasets are generated using Blender software with the Cycles rendering engine, which simulates the complex interactions among light, objects, floors, and shadows. These datasets are suitable for evaluating causal representation learning methods, particularly when addressing computer vision tasks such as distribution shift, domain adaptation, and fairness.