DESOBAv2

DESOBAv2是由上海交通大学教育部人工智能重点实验室创建的大规模真实世界阴影生成数据集，旨在补充现有的小规模数据集DESOBA。该数据集包含28,573个有效实例，通过收集大量户外场景图像，并利用预训练的对象-阴影检测和修复技术构建。数据集的创建过程涉及图像收集、阴影移除和合成图像合成等步骤。DESOBAv2主要应用于图像合成领域，特别是解决前景对象与背景之间的阴影不一致问题，以提高合成图像的真实感。

DESOBAv2 is a large-scale real-world shadow generation dataset created by the Key Laboratory of Artificial Intelligence, Ministry of Education, Shanghai Jiao Tong University, aiming to supplement the existing small-scale dataset DESOBA. It contains 28,573 valid instances, constructed by collecting a large number of outdoor scene images and leveraging pre-trained object-shadow detection and inpainting techniques. The dataset creation process includes steps such as image collection, shadow removal and synthetic image generation. DESOBAv2 is mainly applied in the field of image synthesis, specifically to resolve the inconsistency of shadows between foreground objects and their backgrounds, thereby improving the realism of synthesized images.