自动驾驶行业学习、交流

PAConv: Position Adaptive Convolution with Dynamic Kernel Assembling on Point CloudsMutian Xu1 Runyu Ding1* Hengshuang Zhao2 Xiaojuan Qi1† 1The University of Hong Kong 2University of Oxford*cvpr-2021github：https://github.com/CVMI-Lab/PAConv〇、Introduct

2020-MS-RRFSegNet: Multiscale Regional Relation Feature Segmentation Network for Semantic Segmentation of Urban Scene Point Clouds时间：2020期刊：IEEE Transactions on Geoscience and Remote Sensing作者：Luo, Haifeng、Chen, Chongcheng、Fang, Lina、Khoshelham, Kourosh

PointWeb: Enhancing Local Neighborhood Features for Point Cloud Processingcvpr 2019Hengshuang Zhao地址：https://github.com/hszhao/PointWeb1、创新针对之前的方法（PointNet、dgcnn的局部点云表达弱的问题，提出了一种考虑局部内所有点之间作用的module。2、具体表现主要就是提出一个Adaptive Feature Adjustment（AFA）模块：

Relation-Shape Convolutional Neural Network for Point Cloud Analysis中科院 刘永成CVPR 2019 oral代码：https://github.com/Yochengliu/Relation-Shape-CNN创新为了解决局部点云特征聚合的问题，设计一种点云的卷积，从邻域关系上更新中心点的Feature。核心就是通过邻域点与中心点之间的低层次特征来提取高层次特征，最终得到该邻域点相对于中心点的权重，通过聚合所有的邻域点加权

3D Semantic Segmentation for Large-Scale Scene UnderstandingKiran Akadas and Shankar GangisettyKLE Technological University, Hubballi, Indiaakadask@gmail.com, shankar@kletech.ac.in年份：2020期刊/会议：ACCV代码：https://github.com/KiranAkadas/GRanDNet1、创新基于R

PointConv1、创新1、将连续卷积扩展到三维点云，对中心点的邻域内点云进行卷积2、提出高效的PointConv执行方式（没看懂）3、将卷积推广到反卷积（和PointNet2类似）2、具体体现1、理论推导将连续的卷积核看作是关于一个3D中心点邻域坐标的连续函数其中，W和F均为连续函数，(x,y,z)是参考点的坐标， (δx,δy,δz)表示邻域 G 中的 3D 点的相对坐标，F代表特征，W代表学习到的权重矩阵（也就是卷积核）。为了针对不均匀的问题，增加一个逆密度矩阵S：2、卷积

Lightweight Attention Module for Deep Learning on Classification and Segmentation of 3D Point CloudsYunhao Cui, Yi An, Member, IEEE, Wei Sun, Huosheng Hu, Senior Member, IEEE, and Xueguan Song年份：2020期刊：IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMEN

Exploiting Local and Global Structure for PointCloud Semantic Segmentation with Contextual Point Representations期刊：Nips2019时间：2019code：https://github.com/fly519/ELGS目录Exploiting Local and Global Structure for PointCloud Semantic Segmentation with Cont

FPConv: Learning Local Flattening for Point Convolutioncvpr2020将已有的点云学习方式分为Volumetric-style point convolutionGraph-style point convolutionSurface-style point convolution创新提出一种基于表面的学习方式，通过学习一个权重矩阵，将采样点的的邻域点映射为一个平面，再使用2D CNN进行局部特征聚合。具体实现现将局部点云使用

Multi-Resolution Graph Neural Network for Large-Scale Pointcloud Segmentation1 Introduction2 Related WorkVoxel-based approach:Point-based approach:Graph-based approach:3 Proposed Computational Framework3.1 Clustering algorithm for graph formation3.2 Cluste

PASS3D: Precise and Accelerated Semantic Segmentation for 3D PointCloud一、Abstract二、PASS3D FRAMEWORKA. Stage-1: Accelerated cluster proposalB. Stage-2: Point-wise semantic segmentation1) Data preparation:Coordinate transformationData augmentation2) Learning

RandLA-Net: Efficient Semantic Segmentation of Large-Scale Point CloudsQingyong Hu1, Bo Yang1, Linhai Xie1, Stefano Rosa1, Yulan Guo2,3,*Zhihua Wang1, Niki Trigoni1, Andrew Markham11University of Oxford, 2Sun Yat-sen University, 3National University of

目录PAIRWISE LINKAGE FOR POINT CLOUD SEGMENTATION1. INTRODUCTIONclustersegmentObjectives and Motivation2. PAIRWISE LINKAGECutoffDistance:截止距离的选取Density：密度的计算Pairwise Linkage:建立linkage关系，保存聚类中心点Hierarchical Clustering:分层聚类ClusterMerging:3. P-LINKAGE FOR POINT

2017-Segmentation-Based Classification for 3D PointClouds in the Road EnvironmentBinbin Xiang1, Jian Yao1,∗, Xiaohu Lu1, Li Li1, Renping Xie1, and Jie Li21 Computer Vision and Remote Sensing (CVRS) Lab, School of Remote Sensing and Information Engineerin

A Convolutional Neural Network-Based 3D Semantic Labeling Method for ALS Point CloudsZhishuang Yang, Wanshou Jiang ,*ID , Bo XuID , Quansheng Zhu , San JiangID and Wei HuangState Key Laboratory ...