Removing invalid bounding boxes from datastore

I am trying to use object detector training data create using the image data labeler to train a YOLOv2 model. I keep getting the error:

Error using vision.internal.cnn.validation.checkTrainingBoxes (line 12)

Training data from a read of the input datastore contains invalid bounding boxes. Bounding boxes must be non-empty, fully contained within their

associated image and must have positive width and height. Use datastore transform method and remove invalid bounding boxes.

All entries in the dataset have a corresponding bounding box in the correct format. Does anyone know how to do this as suggested?

% sort data
rng(0);
shuffledIndices = randperm(height(Data));
idx = floor(0.7 * length(shuffledIndices) );
trainingDataTbl = Data(shuffledIndices(1:idx),:);
testDataTbl = Data(shuffledIndices(idx+1:end),:);
% create image datastore
imdsTrain = imageDatastore(trainingDataTbl{:,'imageFilename'});
bldsTrain = boxLabelDatastore(trainingDataTbl(:,'Tip'));
imdsTest = imageDatastore(testDataTbl{:,'imageFilename'});
bldsTest = boxLabelDatastore(testDataTbl(:,'Tip'));
trainingData = combine(imdsTrain,bldsTrain);
testData = combine(imdsTest,bldsTest);

%% Create YOLOv2 network
% Input size for detector.
imageInputSize = [224 224 3];
% define classes
numClasses = width(Data)-1;

% estimate anchor boxes
numAnchors = 7;
trainingDataForEstimation = transform(trainingData,@(data)preprocessData(data,imageInputSize));

[anchorBoxes, meanIoU] = estimateAnchorBoxes(trainingDataForEstimation, numAnchors);

% define feature extraction network
featureExtractionNetwork = resnet50;
featureLayer = 'activation_40_relu';
lgraph = yolov2Layers(imageInputSize,numClasses,anchorBoxes,featureExtractionNetwork,featureLayer);

% train YOLOv2 object detector
options = trainingOptions('sgdm', ...
        'MiniBatchSize', 16, ....
        'InitialLearnRate',1e-3, ...
        'MaxEpochs',12,...
        'Shuffle','every-epoch');
    
% augment data
augmentedTrainingData = transform(trainingData,@augmentData);

% Train the YOLO v2 detector
[detector,info] = trainYOLOv2ObjectDetector(trainingData,lgraph,options);

Answer 1

[ahmed shahin]

please if you try to detect multi object and it is not necessary that each image have all classes

is this causes error?

i will appreciate your help

[wang Wang]

I encountered this problem when detecting multiple objects，does anyone know how to solve it?

Answer 2

[Tunai Marques]

Hi Kyle, I am facing a very similar problem. In my case, I can see exactly what augmented samples/BBs are giving me trouble.

Try running the following code to check the transformed data. One of them should have a funky bounding box (in my case, one of them did not have a BB at all after the transform).

Change "25" by the number of samples you want to check. "trainingData" is the 1x1 TransformedDatastore.

Hope that helps.

augmentedData = cell(25,1);
k = 1; 
while hasdata(trainingData)
    T = read(trainingData);
    I = T{1};
    bbox = T{2};
    augmentedData{k} = insertShape(I,'Rectangle',bbox);
    k = k+1; 
end

figure
montage(augmentedData,'BorderSize',2)

[Madura Meenakshi Ramamoorthi]

Hi Tunai Marques,

How did u remove that funky bounding box from trainingdata.

Answer 3

[Cong Dong Ngoc Minh]

Hi Kyke,

Please check the bounding boxes whether it has '0' in coordinate. Due to MATLAB starts at '1' in the image, so that you can modify from '0' to '1' of the bounding box's coordinate. Scale it if it is larger than the image's size.

Thanks!

[Hamza Afzal]

I think 0 in bboxes doesnot make an error, since also in matlab personal .mat files, there were empty sets