Data Augmentation for images and bounding boxes in Object Detection

Hello all,

I am trying to apply data augmentation on a object detection dataset created with Image Labeler App in MATLAB. As explained in https://fr.mathworks.com/help/deeplearning/ug/object-detection-using-yolo-v2.html, the trainingData which is a table containing imageFilename and the bouding boxes coordinates of definite object classes is augmented, with a defined augmentation function (augmentData found in the same link), using:

augmentedTrainingData = transform(trainingData,@augmentData);

I am trying to apply the same line of code after labeling my own dataset and creating the trainingData from the gTruth saved with:

trainingData = objectDetectorTrainingData(gTruth,'SamplingFactor',1, 'WriteLocation','TrainingData');

Although I followed the same concept explained, when using transform function, I am getting the error below:

Undefined function 'transform' for input arguments of type 'table'.

augmentedTrainingData = transform(trainingData,@augmentData);

How can I apply data augmentation on trainingData? I am using MATLAB R2019a.

Appreciate any kind of help. Thank you in advance !!

Answer 1

[Seth Furman]

objectDetectorTrainingData must be called with 2 outputs in order for the first output to be an image datastore and not a table.

https://www.mathworks.com/help/vision/ref/objectdetectortrainingdata.html

imageDir = fullfile(matlabroot,'toolbox','vision','visiondata','vehicles');
addpath(imageDir);
data = load('vehicleTrainingGroundTruth.mat');
gTruth = data.vehicleTrainingGroundTruth;
vehicleDetector = load('yolov2VehicleDetector.mat');
lgraph = vehicleDetector.lgraph;

% imds is an image datastore.
[imds,~] = objectDetectorTrainingData(gTruth)

imds = 
  ImageDatastore with properties:

                       Files: {
                              '/MATLAB/toolbox/vision/visiondata/vehicles/image_00123.jpg';
                              '/MATLAB/toolbox/vision/visiondata/vehicles/image_00099.jpg';
                              '/MATLAB/toolbox/vision/visiondata/vehicles/image_00174.jpg'
                               ... and 292 more
                              }
                     Folders: {
                              '/MATLAB/toolbox/vision/visiondata/vehicles'
                              }
    AlternateFileSystemRoots: {}
                    ReadSize: 1
                      Labels: {}
      SupportedOutputFormats: ["png"    "jpg"    "jpeg"    "tif"    "tiff"]
         DefaultOutputFormat: "png"
                     ReadFcn: @readDatastoreImage

% trainingDataTable is a table.
trainingDataTable = objectDetectorTrainingData(gTruth)

trainingDataTable = 295×2 table
                            imageFilename                                 vehicle     
    ______________________________________________________________    ________________

    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00123.jpg'}    {2×4 double    }
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00099.jpg'}    {2×4 double    }
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00174.jpg'}    {[ 30 27 97 78]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00294.jpg'}    {[ 93 74 26 25]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00244.jpg'}    {[33 29 103 85]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00281.jpg'}    {[ 90 69 22 19]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00032.jpg'}    {[ 86 69 25 19]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00040.jpg'}    {[ 84 77 38 27]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00022.jpg'}    {[ 83 66 27 24]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00261.jpg'}    {[ 92 70 21 17]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00175.jpg'}    {[56 39 102 78]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00034.jpg'}    {[ 96 70 22 18]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00152.jpg'}    {[ 32 24 98 88]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00092.jpg'}    {2×4 double    }
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00170.jpg'}    {[74 22 111 84]}
    {'/MATLAB/toolbox/vision/visiondata/vehicles/image_00091.jpg'}    {2×4 double    }

[Andrea Daou]

Hello, thank you for your answer.

I am having a problem calling objectDetectorTrainingData with 2 arguments. My gTruth is structured as in the attached capture. How can i get the right format that can be giving to objectDetectorTrainingData ?

Appreciate your help.

[David]

Andrea,

Did you ever figure this out? I am at that juncture myself. Any help would be greatly appreciated.

Thanks in advance,

Dave

[Amanjit Dulai]

You should just be able to call objectDetectorTrainingData directly on an object with the class groundTruth.