downloadDirectory = tempdir;datapath = downloadSydneyUrbanObjects(downloadDirectory);

foldsTrain = 1:3;foldsVal = 4;dsTrain = sydneyUrbanObjectsClassificationDatastore(数据路径，foldsTrain);dsVal = sydneyUrbanObjectsClassificationDatastore(数据路径，foldsVal);

data = read(dsTrain);ptCloud = data{1,1};Label = data{1,2};图pcshow (ptCloud。位置，[0 0 1]，“MarkerSize”现年40岁的“VerticalAxisDir”，“向下”)包含(“X”) ylabel (“Y”) zlabel (“Z”标题(标签)

dsLabelCounts = transform(dsTrain，@(data){data{2} data{1}.Count});labelCounts = readall(dsLabelCounts);标签= vertcat(labelCounts{:，1});counts = vertcat(labelCounts{:，2});

图直方图(标签)

rng(0) [G,classes] = findgroups(标签);numObservations = splitapply(@numel,labels,G);desiredNumObservationsPerClass = max(numObservations);files = splitapply(@(x){randReplicateFiles(x,desiredNumObservationsPerClass)}，dsTrain.Files,G);Files = vertcat(Files {:});dsTrain。文件=文件;

dsTrain。Files = dsTrain.Files(randperm(length(dsTrain.Files)));

dsTrain。MiniBatchSize = 128;dsVal。MiniBatchSize = 128;

dsTrain = transform(dsTrain，@augmentPointCloud);

data = preview(dsTrain);ptCloud = data{1,1};Label = data{1,2};图pcshow (ptCloud。位置，[0 0 1]，“MarkerSize”现年40岁的“VerticalAxisDir”，“向下”)包含(“X”) ylabel (“Y”) zlabel (“Z”标题(标签)

minPointCount = splitapply(@min,counts,G);maxPointCount = splitapply(@max,counts,G);= splitapply(@(x)round(mean(x))，counts,G);stats = table(类，numObservations,minPointCount,maxPointCount,meanPointCount)

numPoints = 1024;dsTrain = transform(dsTrain，@(data)selectPoints(data,numPoints));dsVal = transform(dsVal，@(数据)selectPoints(数据，numPoints));

dsTrain = transform(dsTrain，@preprocessPointCloud);dsVal = transform(dsVal，@preprocessPointCloud);

data = preview(dsTrain);图pcshow(data{1,1}，[0 0 1]，“MarkerSize”现年40岁的“VerticalAxisDir”，“向下”）;包含(“X”) ylabel (“Y”) zlabel (“Z”{1,2})标题(数据)

inputChannelSize = 3;hiddenChannelSize1 = [64128];hiddenChannelSize2 = 256;(参数。在putTransform, state.InputTransform] = initializeTransform(inputChannelSize,hiddenChannelSize1,hiddenChannelSize2);

inputChannelSize = 3;hiddenChannelSize = [64 64];[parameters.SharedMLP1状态。SharedMLP1] = initializeSharedMLP(inputChannelSize,hiddenChannelSize);

inputChannelSize = 64;hiddenChannelSize1 = [64128];hiddenChannelSize2 = 256;(参数。FeatureTransform,状态。FeatureTransform] = initializeTransform(inputChannelSize,hiddenChannelSize,hiddenChannelSize2);

inputChannelSize = 64;hiddenChannelSize = 64;[parameters.SharedMLP2状态。SharedMLP2] = initializeSharedMLP(inputChannelSize,hiddenChannelSize);

inputChannelSize = 64;hiddenChannelSize = [512,256];numClasses = nummel(类);(参数。ClassificationMLP,状态。ClassificationMLP] = initializeClassificationMLP(inputChannelSize,hiddenChannelSize,numClasses);

numEpochs = 40;learnRate = 0.001;l2Regularization = 0.01;learnratdropperiod = 15;learnratdropfactor = 0.5;

gradientDecayFactor = 0.9;squaredGradientDecayFactor = 0.999;

avgGradients = [];avgSquaredGradients = [];

doTraining = false;如果doTraining的末尾列出的configureTrainingProgressPlot函数%示例，初始化训练进度图以显示训练%损失、训练准确率和验证准确率。[lossPlotter, trainAccPlotter,valAccPlotter] = initializtrainingprogressplot;numClasses = nummel(类);迭代= 0;开始= tic;为epoch = 1:numEpochs重置训练和验证数据存储。重置(dsTrain);重置(dsVal);遍历数据集。而hasdata(dsTrain)迭代=迭代+ 1;读取数据。data = read(dsTrain);创建批处理。[XTrain,YTrain] = batchData(data);使用dlfeval和% modelGradients函数。[gradients, loss, state, acc] = dlfeval(@modelGradients,XTrain,YTrain,parameters,state);L2正则化。梯度= dlupdate(@(g,p) g + l2Regularization*p，梯度，参数);使用Adam优化器更新网络参数。。[parameters, avgGradients, avgSquaredGradients] = adamupdate(parameters, gradients, avgSquaredGradients).．.avgGradients avgSquaredGradients迭代.．.learnRate、gradientDecayFactor squaredGradientDecayFactor);更新培训进度。D = duration(0,0,toc(start)，“格式”，“hh: mm: ss”）;标题(lossPlotter。父母,”时代:“+ epoch +，消失:"+字符串(D)) addpoints(lossPlotter，迭代，double(gather(extractdata(loss)))) addpoints(trainAccPlotter，迭代，acc);drawnow结束根据验证数据评估模型。cmat = sparse(numClasses,numClasses);而hasdata (dsVal)获取下一批数据。data = read(dsVal);创建批处理。[XVal,YVal] = batchData(data);计算标签预测。istring = false;YPred = pointnetClassifier(XVal,parameters,state, istrin);选择得分最高的预测作为类别标签% XTest。[~，YValLabel] = max(YVal，[]，1);[~，YPredLabel] = max(YPred，[]，1);收集混淆指标。cmat = aggreateConfusionMetric(cmat,YValLabel,YPredLabel);结束以平均分类准确率更新训练进度图。Acc = sum(diag(cmat))./sum(cmat，“所有”）;addpoints (valAccPlotter迭代,acc);更新学习率。如果mod(epoch, learnratdrop period) == 0 learnRate = learnRate * learnratdrop factor;结束重置训练和验证数据存储。重置(dsTrain);重置(dsVal);结束其他的下载预训练的模型参数、模型状态和验证。%的结果。pretrainedURL =“//www.tianjin-qmedu.com/万博1manbetxsupportfiles/vision/data/pretrainedPointNet.mat”；pretrainedNetwork = fullfile(pwd，“pretrainedPointNet.mat”）;如果~存在(pretrainedNetwork“文件”) disp (“下载预训练的网络(5 MB)…”）;websave (pretrainedNetwork pretrainedURL);结束加载预训练模型。预训练结果=负载(“pretrainedPointNet.mat”）;参数= pretrainedResults.parameters;state = pretrainedResults.state;cmat =预训练结果。cmat;如果可能的话，将模型参数移动到GPU并转换为dlarray。parameters = prepareForPrediction(parameters，@(x)dlarray(toDevice(x,canUseGPU)));如果可能，将模型状态移动到GPU。state = prepareForPrediction(state，@(x)toDevice(x,canUseGPU));结束

Figure chart = confusionchart(cmat,classes);

Acc = sum(diag(cmat))./sum(cmat，“所有”）

ptCloud = pcread(“car.pcd”）;X = preprocessPointCloud(ptCloud);dlX = dlarray(X{1}，“SCSB”）;

YPred = pointnetClassifier(dlX，参数，状态，false);[~， classsidx] = max(YPred，[]，1);

图pcshow (ptCloud。位置，[0 0 1]，“MarkerSize”现年40岁的“VerticalAxisDir”，“向下”标题(类(classIdx))

函数[gradients, loss, state, acc] = modelGradients(X,Y,parameters,state)执行模型函数。istring = true;[YPred,state,dlT] = pointnetClassifier(X,parameters,state, istrin);添加正则化项，确保特征变换矩阵为%近似正交。K = size(dlT,1);B = size(dlT, 4);I = repelem(眼(K)，1,1,1,B);dlI = dlarray(I，“SSCB”）;treg = mse(dlI,pagemtimes(dlT,permute(dlT，[2 1 3 4])));因子= 0.001;计算损失。损失=交叉熵(YPred,Y) +因子*treg;计算关于损失的参数梯度。Gradients = dlgradient(损失，参数);计算训练精度度量。[~，YTest] = max(Y，[]，1);[~，YPred] = max(YPred，[]，1);acc = gather(extractdata(sum(YTest == YPred)./ nummel (YTest)));结束

函数[dlY,state,dlT] = pointnetClassifier(dlX,parameters,state, istrin)调用PointNet编码器。[dlY,state,dlT] = pointnetEncoder(dlX,parameters,state, istrin);调用分类器。p = parameters.ClassificationMLP.Perceptron;s = state.ClassificationMLP.Perceptron;为k = 1:元素个数(p)[海底,s (k)] =感知器(海底,p (k), s (k), isTraining);%如果是训练，则应用概率为0.3的反向dropout。如果istrain概率= 0.3;dropoutScaleFactor = 1 -概率;dropoutMask = (rand(size(dlY)，“喜欢”， dlY) >概率)/ dropoutScaleFactor;ly = ly .*dropoutMask;结束结束state.ClassificationMLP.Perceptron = s;应用最终的全连接和softmax操作。weights = parameters.ClassificationMLP.FC.Weights;偏差=参数。分类mlp . fc .偏差;ly =完全连接(ly，权重，偏差);dlY = softmax(dlY);结束

函数[dlY,state,T] = pointnetEncoder(dlX,parameters,state, istrin)输入变换。(海底,状态。在putTransform] = dataTransform(dlX,parameters.InputTransform,state.InputTransform,isTraining);共享MLP。(海底,state.SharedMLP1。Perceptron] = sharedMLP(dlY,parameters.SharedMLP1.Perceptron,state.SharedMLP1.Perceptron, istrain);%特征变换。(海底,状态。FeatureTransform,T] = dataTransform(dlY,parameters.FeatureTransform,state.FeatureTransform, istrain);共享MLP。(海底,state.SharedMLP2。Perceptron] = sharedMLP(dlY,parameters.SharedMLP2.Perceptron,state.SharedMLP2.Perceptron, istrain);%最大操作。ly = max(ly，[]，1);结束

函数[dlY,state] = sharedMLP(dlX,parameters,state, istring) dlY = dlX;为k = 1: nummel (parameters) [dlY, state(k)] =感知器(dlY,parameters(k)，state(k)， istrain);结束结束

函数[dlY,state] =感知器(dlX,parameters,state, istrin)%卷积。W = parameters. convc . weights;B = parameters. convc . bias;dlconv(dlX,W,B);批处理归一化。训练时更新批归一化状态。offset = parameters.BatchNorm.Offset;scale = parameters.BatchNorm.Scale;trainedMean = state.BatchNorm.TrainedMean;trainedVariance = state.BatchNorm.TrainedVariance;如果istrain [dlY,trainedMean,trainedVariance] = batchnorm(dlY,offset,scale,trainedMean,trainedVariance);%更新状态。state.BatchNorm.TrainedMean = trainedMean;state.BatchNorm.TrainedVariance = trainedVariance;其他的dlY = batchnorm(dlY,offset,scale,trainedMean,trainedVariance);结束% ReLU。dlY = relu(dlY);结束

函数[dlY,state,T] = dataTransform(dlX,parameters,state, istrin)共享MLP。(海底,state.Block1。Perceptron] = sharedMLP(dlX,parameters.Block1.Perceptron,state.Block1.Perceptron, istrion);%最大操作。ly = max(ly，[]，1);共享MLP。(海底,state.Block2。Perceptron] = sharedMLP(dlY,parameters.Block2.Perceptron,state.Block2.Perceptron, istrain);转换网络(T-Net)。将最后一个全连接操作作为W*X应用于%预测变换矩阵T。dlY = extractdata(dlY);dlY =挤压(dlY);% N-by-BT =参数。改造;% K ^ 2-by-B将T重塑成一个方阵。K =√(size(T,1));T =重塑(T,K,K,1，[]);% [k k 1 b]T = T +眼(K);使用批处理矩阵乘法应用于输入dlX。X = extractdata(dlX);% [m 1 k b][C,B] = size(X，[3 4]);X =重塑(X，[]，C,1,B);% [m k 1 b]Y = pagemtimes(X,T);ly = darray (Y，“SCSB”）;结束

函数[params,state] = initializeTransform(inputChannelSize,block1,block2) [params. block1, state.]Block1] = initializeSharedMLP(inputChannelSize, Block1);[params.Block2状态。Block2] = initializeSharedMLP(block1(end)， Block2);%变换矩阵的参数。参数个数。变换= dlarray(0 (inputChannelSize^2,block2(end)));结束

函数[params,state] = initializeSharedMLP(inputChannelSize,hiddenChannelSize) weights = initializeWeightsHe([1 1 inputChannelSize hiddenChannelSize(1)]);偏差= 0 (hiddenChannelSize(1)，1，“单身”）;p.c. . weights = dlarray(weights);p.c. . bias = dlarray(偏差);p. batchnormal . offset = dlarray(0 (hiddenChannelSize(1)，1，“单身”));p. batchnormal . scale = dlarray(ones(hiddenChannelSize(1)，1，“单身”));s.BatchNorm.TrainedMean = 0 (hiddenChannelSize(1)，1，“单身”）;s.BatchNorm.TrainedVariance = ones(hiddenChannelSize(1)，1，“单身”）;params.Perceptron(1) = p;state.Perceptron(1) = s;为k = 2:numel(hiddenChannelSize) weights = initializeWeightsHe([1 1 hiddenChannelSize(k-1) hiddenChannelSize(k)]);偏差= 0 (hiddenChannelSize(k)，1，“单身”）;p.c. . weights = dlarray(weights);p.c. . bias = dlarray(偏差);p. batchnormal . offset = dlarray(0 (hiddenChannelSize(k)，1，“单身”));p. batchnormal . scale = dlarray(ones(hiddenChannelSize(k)，1，“单身”));s.BatchNorm.TrainedMean = 0 (hiddenChannelSize(k)，1，“单身”）;s.BatchNorm.TrainedVariance = ones(hiddenChannelSize(k)，1，“单身”）;params.Perceptron(k) = p;state.Perceptron(k) = s;结束结束

函数[params,state] = initializeClassificationMLP(inputChannelSize,hiddenChannelSize,numClasses) [params,state] = initializeSharedMLP(inputChannelSize,hiddenChannelSize);weights = initializeWeightsGaussian([numClasses hiddenChannelSize(end)]);偏差= 0 (numClasses,1，“单身”）;params.FC.Weights = dlarray(weights);params.FC.Bias = dlarray(bias);结束

函数x = initializeWeightsHe(sz) fanIn = prod(sz(1:3));stddev = sqrt(2/fanIn);X = stddev .* randn(sz);结束

函数x =初始化加权高斯(sz) x = randn(sz，“单身”) .* 0.01;结束

函数data = preprocessPointCloud(data)如果~iscell(data) data = {data};结束numObservations = size(data,1);为i = 1:numObservations刻度点在0到1之间。xlim = data{i,1}.XLimits;ylim = data{i,1}.YLimits;zlim = data{i,1}.ZLimits;xyzMin = [xlim(1) ylim(1) zlim(1)];xyzDiff = [diff(xlim) diff(ylim) diff(zlim)];Data {i,1} = (Data {i,1}.)- xyzMin) ./ xyzDiff;结束结束

函数data = selectPoints(data,numPoints)通过向下抽样或复制选择所需的点数%点云数据。numObservations = size(data,1);为i = 1:numObservations ptCloud = data{i,1};如果ptCloud。计数> numPoints百分比= numPoints/ptCloud.Count;data{i,1} = pcdownsample(ptCloud，“随机”、比例);其他的replicationFactor = ceil(numPoints/ptCloud.Count);ind = repmat(1:ptCloud.Count,1,replicationFactor);data{i,1} = select(ptCloud,ind(1:numPoints));结束结束结束

函数data = augmentPointCloud(data) numObservations = size(data,1);为i = 1:numObservations ptCloud = data{i,1};%围绕“上轴”旋转点云，即该数据集的Z轴。tform = randomAffine3d(.．.“XReflection”,真的,.．.“YReflection”,真的,.．.“旋转”, @randomRotationAboutZ);ptCloud = pctransform(ptCloud,tform);%随机掉分30%。如果0.5 ptCloud = pcdownsample(ptCloud，“随机”, 0.3);结束如果兰特> 0.5%用高斯噪声抖动点位置，均值为0和通过创建一个随机位移场，使标准偏差为0.02。D = 0.02 * randn(size(ptCloud.Location));ptCloud = pctransform(ptCloud,D);结束data{i,1} = ptCloud;结束结束函数[rotationAxis,theta] = randomRotationAboutZ() rotationAxis = [0 0 1];= 2* π *rand;结束

函数cmat = aggreateConfusionMetric(cmat,YTest,YPred) YTest = gather(extractdata(YTest));YPred = gather(extractdata(YPred));[m,n] = size(cmat);cmat = cmat + full(稀疏(YTest,YPred,1,m,n));结束

函数[plotter,trainAccPlotter,valAccPlotter] = initializeTrainingProgressPlot()绘制损失、训练准确率和验证准确率。数字%损失图Subplot (2,1,1) plot = animatedline;包含(“迭代”) ylabel (“损失”）%准确度图subplot(2,1,2) trainAccPlotter = animatedline(“颜色”，“b”）;valAccPlotter = animatedline(“颜色”，‘g’）;传奇(“训练的准确性”，“验证准确性”，“位置”，“西北”）;包含(“迭代”) ylabel (“准确性”）结束

函数randReplicateFiles(files,numDesired) n = numel(files);ind = randi(n,numDesired,1);文件=文件(ind);结束

函数datapath = downloadSydneyUrbanObjects(dataLoc)如果nargin == 0 dataLoc = pwd;结束dataLoc = string(dataLoc);url =“http://www.acfr.usyd.edu.au/papers/data/”；name =“sydney-urban-objects-dataset.tar.gz”；datapath = fullfile(dataLoc，“sydney-urban-objects-dataset”）;如果~存在(datapath公司,“dir”) disp (“正在下载悉尼城市对象数据集……”）;解压(url +名字,dataLoc);结束结束

函数[dlX,海底]= batchData(数据)X =猫(4,数据{:1});标签= cat(1,data{:，2});Y = oneHotEncode(标签);将数据转换为单个数据进行处理。X =单(X);Y =单(Y);尽可能将数据移动到GPU。如果canUseGPU X = gpuArray(X);Y = gpuArray(Y);结束返回X和Y作为dlarray对象。dlX = dlarray(X，“SCSB”）;ly = darray (Y，“CB”）;结束

函数Y = oneHotEncode(labels) numObservations = numel(labels);numCategories = nummel(类别(标签));sz = [numCategories, numObservations];Y = 0 (sz，“单一”）;标签=标签';idx = sub2ind(sz, int32(labels)， 1:numObservations);Y(idx) = 1;结束

函数p = prepareForPrediction(p,fcn)为i = 1:元素个数(p) p (i) = structfun (@ (x)调用(fcn x)、p (i),“UniformOutput”, 0);结束函数数据=调用(fcn, Data)如果isstruct(data) data = prepareForPrediction(data,fcn);其他的数据= fcn(数据);结束结束结束将数据移动到GPU。函数x = toDevice(x,useGPU)如果useGPU x = gpuArray(x);结束结束