LUCES

从阴影信息对对象进行三维重建是计算机视觉中的一项艰巨任务。由于面对光度立体问题的大多数方法都使用简化的远场假设，因此现实世界中的场景本质上具有更复杂的物理效果，需要处理这些物理效果以准确地重建3D形状。已经提出了越来越多的方法来解决假定点光源在目标对象附近时的问题。为了了解处理这种近场场景的方法的能力，迄今为止的文献已经使用了合成渲染的光度图像或最小且非常定制的现实世界数据。为了填补评估近场光度立体方法的空白，我们介绍了LUCES的第一个真实世界的 “近场点光源光度立体数据集”，该数据集涉及14个不同材料的物体。已使用52个led照亮距离相机10至30厘米的每个物体。与原始图像一起，为了评估3D重建，数据集包括法线图和深度图，用于比较检索到的3D几何形状的不同特征。此外，我们评估了拟议数据集上最新的近场光度立体算法的性能，以评估有关实际近距离效应和物体材料的最新方法。

3D object reconstruction from shadow information is a challenging task in computer vision. Since most methods for the photometric stereo problem rely on simplified far-field assumptions, real-world scenes inherently exhibit more complex physical effects that must be addressed for accurate 3D shape reconstruction. A growing number of methods have been proposed to address the scenario where point light sources are placed in close proximity to target objects. To understand the capabilities of methods that handle such near-field scenes, existing literature has relied on either synthetically rendered photometric images or minimal, highly customized real-world datasets. To fill the gap in evaluating near-field photometric stereo methods, we introduce the first real-world "Near-Field Point-Light Photometric Stereo Dataset" from LUCES, which features 14 objects with distinct materials. Each object is illuminated by 52 LEDs positioned 10 to 30 centimeters away from the camera. Alongside the raw images, the dataset includes normal maps and depth maps to enable evaluation of 3D reconstruction, allowing comparison of different characteristics of the retrieved 3D geometry. Additionally, we evaluate the performance of state-of-the-art near-field photometric stereo algorithms on the proposed dataset to assess how current methods handle real-world proximity effects and object materials.

搜集汇总

背景与挑战

背景概述

LUCES是一个真实世界的近场点光源光度立体数据集，包含14个不同材料的物体，每个物体由52个LED灯照亮，提供原始图像、法线图和深度图，用于3D重建评估。该数据集由剑桥大学·东芝欧洲剑桥研究实验室于2021年发布，支持近场光度立体方法的研究。