基于镜面对称引导的单幅图像三维重建  

3D Point Cloud Reconstruction from Single Image Guided by Mirror Symmetry

作　　者：路志青 付燕平 Lu Zhiqing;Fu Yanping(School of Computer Science and Technology,Anhui University,Hefei 230601)

机构地区：[1]安徽大学计算机科学与技术学院,合肥230601

出　　处：《现代计算机》2022年第5期91-97,共7页Modern Computer

基　　金：国家自然科学基金(62076003)。

摘　　要：从单幅图像中恢复三维结构的一个难点问题就是当图像中物体存在自遮挡时,会使得推断的三维结构具有明显的二义性。受到自然界的对象以及人造物体内在的镜面对称性结构的启发,我们通过概率预测来为每一个重建物体自身的镜像对称平面,并利用这个镜像对称平面作为约束来引导物体以自对称的方式生成三维结构,并有效地恢复图像中物体自遮挡部分的三维结构。本文在ShapeNet和Pix3D数据集上进行大量实验,并进行定性和定量分析,实验结果证明了我们提出的基于对称引导的三维重建网络能够有效地提高点云生成的完整度和精度。One of the difficulties in recovering 3D structure from a single image is that when there is self-occlusion in the image, the inferred 3D structure will have obvious ambiguity. Inspired by the mirror symmetry structure of natural objects and manmade objects, we reconstruct the mirror symmetry plane of each object through probability prediction, and use this mirror symmetry plane as a constraint to guide the object to generate three-dimensional structure by self-symmetry, and effectively restore the three-dimensional structure of the self-occluded part of the image. In this paper, a large number of experiments are carried out on ShapeNet and Pix3D datasets, and qualitative and quantitative analysis is carried out. The experimental results show that our proposed 3D reconstruction network based on symmetry guidance can effectively improve the integrity and accuracy of point cloud generation.

关 键 词：单幅图像3D重建 镜面对称 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]