No pixel label data found - using the pixelLabelTrainingData function...

Hello!

I am working an image semantic segmentation project; I am using the pixelLabelTrainingData function to generate two datastores - one image datastore and the second a pixel datastore.

These are the steps I'm taking:

1. Using the Image Labeler app to manually define pixel labels on a few images; output from this step is a gTruth object
2. Use the gTruth object as input into pixelLabelTrainingData; output are two datastores - imds and pxdsTruth
3. Here I will partition data, then create, train network.

BUT - I am stuck at step 2!

Error using pixelLabelTrainingData>sampleDataSource (line 175)

Unrecognized function or variable 'Error'.

No pixel label data found in any of the input ground truth objects.

Error in pixelLabelTrainingData (line 124)
sampleIndices = sampleDataSource(gTruth, labelDefs, samplingFactor);

Attached is the gTruth object I am using.... any help would be greatly appreciated!!

[Walter Roberson]

My guess would be that at line 175 of pixelLabelTrainingData that there is a call to Error() that should have been a call to error()

The error that it is trying to show would be

No pixel label data found in any of the input ground truth objects.

and the corrected line would be

    error(message('vision:pixelLabelTrainingData:NoPixelLabelsInGroundTruth'));

This is line 193 for R2021b but could well be earlier for your R2020a.

[Emmanuel J Rodriguez]

Thank you for this. Any suggestion as to how I can use this information to solve my issue?

The error indicates that there is...

NoPixelLabelsInGroundTruth

... but the attached gTruthTraining.mat file DOES contain PixelLabel. Thanks again.

Answer 1

[yanqi liu]

yes，sir，when load the mat file，all field is empty， please check，such as

a = load('gTruthSeq.mat');

Warning: The data source points to a directory that cannot be found.
'C:\CFP\building'
Update the DataSource using changeFilePaths method.

a.gTruth.LabelDefinitions

ans = 3×5 table
        Name           Type       PixelLabelID     Group      Description
    ____________    __________    ____________    ________    ___________

    {'Sky'     }    PixelLabel    {[       1]}    {'None'}    {0×0 char} 
    {'Lawn'    }    PixelLabel    {[       2]}    {'None'}    {0×0 char} 
    {'Building'}    Rectangle     {0×0 double}    {'None'}    {0×0 char} 

[Emmanuel J Rodriguez]

Thank you for the response, I followed your advice and this is what I get:

>> load('gTruthTraining.mat')
>> a = load('gTruthTraining.mat');
>> a.gTruth.LabelDefinitions

ans =

  2×6 table

         Name             Type                LabelColor              PixelLabelID             Group              Description
    _______________    __________    _____________________________    ____________    ________________________    ___________

    {'Part'       }    PixelLabel    0.46667    0.67451    0.18824       {[1]}        {'AM_build_environment'}    {0×0 char} 
    {'Build_plate'}    PixelLabel    0.92941    0.69412    0.12549       {[2]}        {'AM_build_environment'}    {0×0 char} 

Doesn't this table indicate that there IS pixel label data? - I don't quite understand what I am missing... :(

[yanqi liu]

yes，sir，please use

a.gTruth.DataSource

get the label folder，such as

dataDir = fullfile(toolboxdir('vision'),'visiondata');
pxDir = fullfile(dataDir,'buildingPixelLabels');
winopen(pxDir);

then check the label file

[Emmanuel J Rodriguez]

Thank you. But what exactly do you mean by "then check the label file"? What exactly am I checking?

[yanqi liu]

yes，sir，may be we check the label file，just make sure the label is valid，such as value 1--->label 1，value 2--->lablel 2

[Emmanuel J Rodriguez]

Okay:

View exported pixel label data

unlabeledImg = readimage(imdsTrain,1); % Unlabeled
imshow(unlabeledImg);

Define categories for each pixel value in labeledImg.

labelIDs = [gTruth.LabelDefinitions.PixelLabelID{1},gTruth.LabelDefinitions.PixelLabelID{2}];
labelCats = ["Part", "Build_plate"];

Construct a categorical matrix using the image and the definitions.

buildLabelCats = categorical(pixelLabeledImg,labelIDs,labelCats);

Display the categories overlaid on the image.

figure;
imshow(labeloverlay(unlabeledImg,buildLabelCats))

Doesn't all of this look "normal"? - I don't understand why the below function is not working?

[imds,pxdsTruth] = pixelLabelTrainingData(gTruth,'SamplingFactor',5,'WriteLocation','TrainingData\');

Error using pixelLabelTrainingData>sampleDataSource (line 175)

No pixel label data found in any of the input ground truth objects.

Error in pixelLabelTrainingData (line 124)

sampleIndices = sampleDataSource(gTruth, labelDefs, samplingFactor);

[yanqi liu]

yes，sir，i think the gTruth.LabelData ， labelDefs may be empty，please debug，such as

Answer 2

[Chong Kian Chee]

Hi, may I know if the issue is resolved? I am facing the same issue here!

%%%Updated on 8 January 2022%%

I hope this is stil helpful, the problem is due to the rectangle ROI which does not have PixelLabelID. This might due to the non-pixel labelling ROI that you had used to label ur dataset in the ImageLabeller. To convert this rectangle ROI into a pixel label with a pixel ID, refer to the link below:

https://www.mathworks.com/help/images/ref/polytoblockedimage.html#d123e230604

https://www.mathworks.com/help/images/convert-image-labeler-polygons-to-labeled-blocked-images-for-semantic-segmentation.html

https://www.mathworks.com/help/vision/ug/label-objects-using-polygons.html

Hope this resolve your issue , cheers :)

Answer 3

[Jose Agustin Barrachina]

Same issue here... Any insight? What is exactly the issue? My LabelData looks like the attached image. I only labeled one of the 3 images (because I only care about that one). Is that the issue?