LunarStereo
收藏资源简介:
LunarStereo数据集是首个公开的、高保真的月球立体图像数据集,专为基于深度学习的3D重建设计。该数据集包括通过光线追踪生成的模拟立体对,基于高分辨率数字高程模型(DEM)进行渲染,并使用精确的反射率模型(BRDF)、逼真的太阳照明和变化的相机轨迹。它涵盖了月球南极周围的各种高度、光照条件和视角,为3D重建任务提供了物理基础的监督。通过使用这个数据集对MASt3R模型进行微调,展示了其在月球领域中的成功应用,显著提高了3D几何估计的准确性和相对准确性。
The LunarStereo Dataset is the first publicly available, high-fidelity lunar stereo image dataset designed specifically for deep learning-based 3D reconstruction. This dataset includes simulated stereo pairs generated via ray tracing, rendered based on high-resolution digital elevation models (DEMs), and employs precise bidirectional reflectance distribution function (BRDF) models, realistic solar illumination, and varying camera trajectories. It covers diverse elevations, lighting conditions, and viewing perspectives around the lunar south pole, providing physically grounded supervision for 3D reconstruction tasks. Fine-tuning the MASt3R model with this dataset has demonstrated its successful application in lunar scenarios, significantly improving the accuracy and relative accuracy of 3D geometry estimation.

- 1通过IRIT, Toulouse INP, Universite de Toulouse, France; Airbus Defence and Space; ESA (European Space Agency) · 2025年



