基于平面补丁的自适应八叉树三维图像重建  

Adaptive octree 3D image reconstruction based on plane patch

作　　者：姚程 马彩文[1,2] YAO Cheng;MA Caiwen(Xi'an Institute of Optics and Precision Mechanics,Chinese Academy of Sciences,Xi'an 710119,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院西安光学精密机械研究所,陕西西安710119 [2]中国科学院大学,北京100049

出　　处：《光学精密工程》2022年第9期1113-1122,共10页Optics and Precision Engineering

基　　金：吉林省自然科学基金资助项目(No.20150520059JH)。

摘　　要：提出了一种基于平面补丁的自适应八叉树卷积神经网络(Octree Convolutional Neural Networks,O-CNN),用于进行有效的三维形状编码和解码。不同于基于体素或基于八叉树的卷积神经网络(Convolutional Neural Networks,CNN),以相同的分辨率表示具有体素的三维形状,O-CNN可自适应地表示具有不同层次的八叉树节点的三维形状,并使用平面补丁对每个八叉树节点内的三维形状进行建模。依据这种自适应表示设计了一种用于编码和解码三维形状的自适应O-CNN编码器和解码器。自适应O-CNN编码器将平面补丁法线和位移作为输入,仅在每个级别的八叉树节点上执行三维卷积操作,而自适应O-CNN解码器则推断每个层次的八叉树节点的形状占有率和细分状态,并估计每个最佳叶八叉树节点的平面法线和位移。通过对单个图像的形状预测验证了自适应O-CNN的生成任务的效率和有效性,倒角距离误差为0.274,低于OctGen的倒角距离误差0.294,取得了更好的重建效果。作为3D形状分析和生成的通用框架,基于平面补丁的自适应O-CNN不仅减少了内存和计算成本,而且比现有的3D-CNN方法具有更好的形状生成能力。In this study,an adaptive octree convolutional neural network based on plane patches is proposed for effective 3D shape encoding and decoding.Unlike volume-based or octree-based convolutional neural network(CNN)methods,which represent 3D shapes with the same voxel resolution,the proposed method can use planes and adaptively represent the 3D shapes of octree nodes with different levels.The patch models the 3D shape within each octree node,whereby the patch-based adaptive representation is utilized in the proposed adaptive patch octree convolutional neural network(O-CNN)encoder and decoder for the encoding and decoding of 3D shapes.The adaptive patch O-CNN encoder takes the plane patch normal and displacement as input and performs three-dimensional convolution on the octree nodes of each level,whereas the adaptive patch O-CNN decoder infers each level.The shape occupancy rate and subdivision state of the octree node as well as the best plane normal and displacement of each leaf octree node are estimated.As a general framework for 3D shape analysis and generation,adaptive patch O-CNN not only reduces memory and computational costs but also exhibits better shape generation capabilities than existing 3D-CNN methods.Shape prediction is performed using a single image to verify the efficiency and effectiveness of the generation task of the adaptive O-CNN.The chamfer distance error is 0.274,which is lower than that of OctGen(0.294),resulting in a better reconstruction effect.

关 键 词：计算机视觉 三维重建 卷积神经网络 神经网络 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]