Main Content

Calculate Segmentation Metrics in Block-Based Workflow

This example shows how to calculate the semantic segmentation confusion matrix for individual blocks in a categorical bigimage object, then calculate global and block segmentation metrics.

Load a pretrained network that performs binary segmentation of triangles against a background.


The triangleImages data set has 100 test images with ground truth labels. Define the location of the data set.

dataSetDir = fullfile(toolboxdir('vision'),'visiondata','triangleImages');

Define the location of the test images.

testImagesDir = fullfile(dataSetDir,'testImages');

Read three test images. Resize each image by a factor of four, convert it to data type double, then create a bigimage object. A bigimage supports block-based image processing workflows.

numImages = 3;
for idx = 1:numImages
    im = imread(fullfile(testImagesDir,['image_' '00' num2str(idx) '.jpg']));
    im = imresize(im,4);
    testImages(idx) = bigimage(im);

Display the first test image.


Define the location of the ground truth labels.

testLabelsDir = fullfile(dataSetDir,'testLabels');

Define the class names and their associated label IDs.

classNames = ["triangle","background"];
labelIDs   = [255 0];

Read in the ground truth labels for each test image. Create a categorical bigimage object from the ground truth label.

for idx = 1:numImages
    gtLabel = imread(fullfile(testLabelsDir,['labeled_image_' '00' num2str(idx) '.png']));
    gtLabel = imresize(gtLabel,4,'nearest');
    groundTruthImages(idx) = bigimage(gtLabel, ...

Display the first ground truth image.


For each test image, use the apply function to process each block. The apply function performs the operations specified by the helper function segmentAndCalculateBlockMetrics, which is defined at the end of this example. The function performs semantic segmentation of each block and calculates the confusion matrix between the predicted and ground truth labels.

blockSize = [32 32];
datasetConfMat = table();
for idx = 1:numImages
    [segmentedImages(idx),blockConfMatOneImage] = apply(testImages(idx),1, ...
        @(block,labeledImageBlock,blockInfo) segmentAndCalculateBlockMetrics(block,labeledImageBlock,net,blockInfo), ...
        groundTruthImages(idx),'PadPartialBlocks',true, ...
    % Add an image number corresponding to the block results for each image
    blockConfMatOneImage.ImageNumber = idx.*ones(height(blockConfMatOneImage),1);
    datasetConfMat = [datasetConfMat;blockConfMatOneImage];

Display the first segmented image.


Evaluate the data set metrics and block metrics for the segmentation.

[metrics,blockMetrics] = evaluateSemanticSegmentation(datasetConfMat,classNames,'Metrics','all');
Evaluating semantic segmentation results
* Selected metrics: global accuracy, class accuracy, IoU, weighted IoU.
* Processed 3 images.
* Finalizing... Done.
* Data set metrics:

    GlobalAccuracy    MeanAccuracy    MeanIoU    WeightedIoU
    ______________    ____________    _______    ___________

       0.95428          0.82739       0.69927      0.92533  

Calculate the Jaccard score for all images.

jaccardSimilarity = metrics.ImageMetrics.MeanIoU
jaccardSimilarity = 3×1


Supporting Function

The segmentAndCalculateBlockMetrics function performs semantic segmentation of a single block then calculates the confusion matrix of the predicted and ground truth labels.

function [outputLabeledImageBlock,blockConfMatPerBlock] = segmentAndCalculateBlockMetrics(block,labeledImageBlock,net,blockInfo)
    outputLabeledImageBlock = semanticseg(block,net);
    confusionMatrix = segmentationConfusionMatrix(outputLabeledImageBlock,labeledImageBlock);
    % blockConfMatPerBlock is a struct with confusion matrices and
    % blockInfo. Use the struct with evaluateSemanticSegmentation to
    % calculate metrics and aggregate block-based results.
    blockConfMatPerBlock.ConfusionMatrix = confusionMatrix;
    blockConfMatPerBlock.BlockInfo = blockInfo;

See Also

| | | |

Related Examples

More About