Implicit surface reconstruction based on object single view

Journal Information

J. Hefei Univ. Tech. (Nat. Sci.), May.2024, Vol.47 No.5: 642-648

DOI: 10.3969/j.issn.1003-5060.2024.05.010

Authors

XING Yan $ ^{1} $, NIU Saihu $ ^{1} $, HONG Peilin $ ^{2} $, TAN Jieqing $ ^{1} $

(1. School of Mathematics, Hefei University of Technology, Hefei 230601, China; 2. School of Medical Information Engineering, Anhui University of Chinese Medicine, Hefei 230012, China)

Abstract and Keywords

Abstract: The three-dimensional(3D) reconstruction method based on implicit surface provides a good trade-off in terms of fidelity, flexibility and compression ability. In this paper, the implicit surface network is used to learn the 3D surface of an object shape. Firstly, the global feature is extracted from the image by using visual geometry group-16(VGG-16) network, and the local feature is obtained from the VGG-16 network for each sampling point in the modeling space. Then, each sampling point is encoded by a multi-layer perceptron(MLP) to obtain the point feature. Furthermore, the global feature and the local feature are concatenated with the point feature respectively and fed into the two decoders respectively to obtain the magnitude and sign of the signed distance function(SDF) of the sampling point in the implicit field. Finally, the implicit surface of the object is obtained. The proposed method has performed 3D object reconstruction task on ShapeNet datasets with both qualitative and quantitative evaluations superior to state-of-the-art methods, especially for complex topological objects with holes and thin structures.

Keywords: three-dimensional(3D) reconstruction; global feature; local feature; deep learning; signed distance function(SDF)

作者信息

邢燕 $ ^{1} $，牛赛虎 $ ^{1} $，洪沛霖 $ ^{2} $，檀结庆 $ ^{1} $

(1. 合肥工业大学数学学院, 安徽合肥 230601; 2. 安徽中医药大学医药信息工程学院, 安徽合肥 230012)

摘要和关键词

摘要: 基于隐式曲面的三维重建方法在保真度、灵活性和压缩能力方面提供了良好的权衡。文章利用隐式曲面网络学习物体形状的三维表面，首先利用视觉几何群（visual geometry group-16，VGG-16）网络从图像中提取全局特征，对建模空间中的每个采样点从 VGG-16 网络中获取局部特征；其次对每个采样点利用多层感知器（multi-layer perceptron，MLP）进行位置编码得到点特征；然后将全局特征和局部特征分别与点特征串联起来送入 2 个解码器中，获得隐式场中采样点的符号距离函数（signed distance function，SDF）的大小与符号，并最终得到物体的隐式曲面。文中所提出的方法应用于 ShapeNet 数据集上进行三维对象重建任务，定性和定量评估均优于现有方法，特别是对于具有孔洞和薄结构的复杂拓扑物体。

关键词: 三维重建；全局特征；局部特征；深度学习；符号距离函数（SDF）

Funding

国家自然科学基金资助项目(62172135); 合肥工业大学校级教研资助项目(KCSZ2022034)和安徽中医药大学教研重点资助项目(2020xjjy_zd005)

合肥工业大学学报(自然科学版)

Navigation