Pano3D
收藏资源简介:
Pano3D数据集是由希腊研究和科技中心与马德里理工大学合作创建,专注于计算机视觉任务中的单目深度估计。该数据集通过合成技术生成,包含高质量的球形全景图和深度图,旨在解决模型在不受控制的野外数据测试中的分布偏移问题。数据集通过分解为三个不同的分布偏移(协变量、先验和概念)来评估模型的性能。Pano3D数据集的应用领域主要集中在提高深度估计模型的泛化能力和鲁棒性,尤其是在复杂的数据收集过程中。
The Pano3D dataset is developed jointly by the Greek Research and Technology Center and the Polytechnic University of Madrid, focusing on monocular depth estimation tasks in computer vision. Generated via synthetic techniques, this dataset contains high-quality spherical panoramic images and depth maps, aiming to address the distribution shift problem of models when tested on uncontrolled real-world outdoor data. It evaluates model performance by decomposing distribution shifts into three distinct categories: covariate shift, prior shift, and conceptual shift. The primary applications of the Pano3D dataset center on enhancing the generalization ability and robustness of depth estimation models, especially in complex data collection scenarios.




