人工智能培训

搜索

深度学习论文:用于形状分类的规范和紧凑点云表示(Canonical and Compact Point Cloud Representation for Sha

[复制链接]
kjdshfjsfgsfg 发表于 2018-9-14 08:56:49 | 显示全部楼层 |阅读模式
kjdshfjsfgsfg 2018-9-14 08:56:49 40 0 显示全部楼层
深度学习论文:用于形状分类的规范和紧凑点云表示(Canonical and Compact Point Cloud Representation for Shape  Classification)我们提出了一种新颖的紧凑点云表示,其本质上不随尺度,坐标变化和点置换而变化。关键思想是以无监督的方式将围绕单个形状的距离场参数化为唯一的,规范的和紧凑的矢量。我们首先使用奇异值分解将adistance字段投影到$ 4 $ D规范空间。 Wethen为每个实例训练一个神经网络,将其距离场非线性地嵌入到网络参数中。我们采用具有ReLU激活单元的无偏差Extreme LearningMachine(ELM),其在层之间具有比例因子可交换属性。我们通过将它们用于$ 3 $ D数据集中的分类形式来演示实例,形状嵌入式网络参数的描述性。我们基于学习的表示需要最小化和简单的神经网络,其中先前的方法需要许多表示来处理坐标变化和点置换。
We present a novel compact point cloud representation that is inherentlyinvariant to scale, coordinate change and point permutation.The key idea is toparametrize a distance field around an individual shape into a unique,canonical, and compact vector in an unsupervised manner.We firstly project adistance field to a $4$D canonical space using singular value decomposition.Wethen train a neural network for each instance to non-linearly embed itsdistance field into network parameters.We employ a bias-free Extreme LearningMachine (ELM) with ReLU activation units, which has scale-factor commutativeproperty between layers.We demonstrate the descriptiveness of theinstance-wise, shape-embedded network parameters by using them to classifyshapes in $3$D datasets.Our learning-based representation requires minimalaugmentation and simple neural networks, where previous approaches demandnumerous representations to handle coordinate change and point permutation.深度学习论文:用于形状分类的规范和紧凑点云表示(Canonical and Compact Point Cloud Representation for Shape  Classification) zB59RHDzmzLbBl97.jpg
URL地址:https://arxiv.org/abs/1809.04820     ----pdf下载地址:https://arxiv.org/pdf/1809.04820    ----深度学习论文:用于形状分类的规范和紧凑点云表示(Canonical and Compact Point Cloud Representation for Shape  Classification)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

kjdshfjsfgsfg当前离线
新手上路

查看:40 | 回复:0

快速回复 返回顶部 返回列表