Monocular 3D reconstruction is to reconstruct the shape of object and its other detailed information from a single RGB image. In 3D reconstruction, polygon mesh is the most prevalent expression form obtained from deep learning models, with detailed surface information and low computational cost. However, some state-of-the-art works fail to generate well-structured meshes, these meshes have two severe problems which we call Vertices Clustering and Illegal Twist. By delving into the mesh deformation procedure, we pinpoint the inadequate usage of Chamfer Distance(CD) metric in deep learning model. In this paper, we initially demonstrate the problems resulting from CD with visual examples and quantitative analyses. To solve these problems, we propose a fine-grained reconstruction method CD$^2$ with Chamfer distance adopted twice to perform a plausible and adaptive deformation. Extensive experiments on two 3D datasets and the comparison of our newly proposed mesh quality metrics demonstrate that our CD$^2$ outperforms others by generating better-structured meshes.
翻译:单体 3D 重建是用一个 RGB 图像重建对象形状和其他详细信息。 在 3D 重建中, 多边形网状是深层学习模型中最常用的表达形式, 其表层信息详细, 计算成本低。 然而, 一些最先进的工程无法产生结构完善的 meshes, 这些网状有两个严重的问题, 我们称之为 Vertics 集成和非法的 Twist 。 通过对网状变形程序进行剖析, 我们发现在深层学习模型中Chamfer Trace(CD) 衡量标准使用不足。 在本文中, 我们最初用视觉示例和定量分析来展示了 CD 产生的问题。 为了解决这些问题, 我们提出了一种精细的重建方法 CD$2, 使用Chamfer 的距离两次来进行貌似和适应性变形。 有关两个3D 数据集的广泛实验和我们新提议的网状质量指标的比较表明, 我们的CD$2 $ 超越了其它标准, 产生结构更好的 meshes。