How can i modify step 3 section of code, to visualize several objects in the image?

%% Step 1: Define Paths and Initialize Data

% Define paths to train, test, and valid folders

baseFolder = 'C:\Users\grace\Downloads\dataset';

folders = {'train', 'test', 'valid'};

% Initialize a structure to store data tables for each set

datasetTables = struct();

% Iterate over each dataset folder (train, test, valid)

for f = 1:numel(folders)

datasetFolder = fullfile(baseFolder, folders{f});

% Step 1: Load class names from _darknet.labels (only need to load once)

if f == 1

labelFilePath = fullfile(datasetFolder, '_darknet.labels');

classNames = readlines(labelFilePath); % Read class names into a string array

disp('Class names in _darknet.labels:');

disp(classNames);

end

%AB

classNames = replace(classNames, " ", "_");

% Step 2: Create an ImageDatastore for images

imageFiles = dir(fullfile(datasetFolder, '*.jpg')); % Get all .jpg files

imagePaths = fullfile(datasetFolder, {imageFiles.name}); % Full paths to images

imds = imageDatastore(imagePaths); % Create ImageDatastore

% Step 3: Load annotations from .txt files

annotationFiles = dir(fullfile(datasetFolder, '*.txt')); % Get all .txt files

annotationFiles = annotationFiles(~strcmp({annotationFiles.name}, '_darknet.labels')); % Exclude _darknet.labels

annotationPaths = fullfile(datasetFolder, {annotationFiles.name}); % Full paths to annotations

% Initialize a cell array to store annotations

annotations = cell(numel(annotationPaths), 1);

% Read annotations from each .txt file

for i = 1:numel(annotationPaths)

fileID = fopen(annotationPaths{i}, 'r');

annotations{i} = fscanf(fileID, '%c'); % Read the entire file as text

fclose(fileID);

end

% Step 4: Verify alignment between images and annotations

[~, imageNames] = cellfun(@fileparts, imagePaths, 'UniformOutput', false);

[~, annotationNames] = cellfun(@fileparts, annotationPaths, 'UniformOutput', false);

if ~isequal(imageNames, annotationNames)

error('Image and annotation files in %s do not match. Check file names.', folders{f});

end

% Step 5: Prepare data table for this dataset

firstImage = imread(imagePaths{1});

imageSize = size(firstImage); % imageSize = [height, width, channels]

imageSize = imageSize(1:2); % Extract only height and width

% Initialize a table to store image file names, bounding boxes, and labels

% @ABAssume classNames is an Mx1 string array

M = numel(classNames);

% Define variable names: 'imageFilename' followed by class names

variableNames = ["imageFilename", classNames']; % Convert to row for concatenation

% Define variable types: first column as 'string', the rest as 'cell'

variableTypes = ["string", repmat("cell", 1, M)];

% Create the table

data = table('Size', [numel(imagePaths), M+1], ...

'VariableTypes', cellstr(variableTypes), ...

'VariableNames', cellstr(variableNames));

% @AB OLD data = table('Size', [numel(imagePaths), 3], ...

% 'VariableTypes', {'string', 'cell', 'cell'}, ... % Ensure 'boxes' and 'labels' are cell arrays

% 'VariableNames', {'imageFilename', 'boxes', 'labels'});

% Parse annotations and populate the table

for i = 1:numel(imagePaths)

annotation = annotations{i};

annotationLines = splitlines(annotation); % Split into individual lines

% @AB Initialize arrays to store bounding boxes and labels for this image

%boxes = [];

%labels = [];

boxes = cell(1,M);

%labels = cell(1,M);

% Parse each line of the annotation

for j = 1:numel(annotationLines)

if isempty(annotationLines{j})

continue; % Skip empty lines

end

% Split the line into components (YOLO format: class_id x_center y_center width height)

components = str2double(strsplit(annotationLines{j}));

class_id = components(1) + 1; % YOLO class IDs start from 0, MATLAB starts from 1

x_center = components(2);

y_center = components(3);

width = components(4);

height = components(5);

% Convert YOLO format to [x_min y_min width height]

x_min = (x_center - width / 2) * imageSize(2); % imageSize(2) = width

y_min = (y_center - height / 2) * imageSize(1); % imageSize(1) = height

boxes{class_id} = [boxes{class_id}; x_min y_min width * imageSize(2) height * imageSize(1)];

% Add the corresponding label

% labels{class} = [labels; classNames(class_id)];

end

% Debug: Display labels for the current image

% disp(['Labels for image ', num2str(i), ':']);

% disp(labels);

% Add the data to the table

data.imageFilename(i) = imagePaths{i};

%data.boxes{i} = {boxes}; % Wrap in a cell array

for jj=1:M

data.(classNames{jj}){i}=boxes{jj};

end

%data.boxes{i} = {boxes}; % Wrap in a cell array

%data.labels{i} = {categorical(labels, classNames)}; % Wrap in a cell array

end

% Store the data table in the datasetTables structure

datasetTables.(folders{f}) = data;

% Display a preview of the first few rows

fprintf('Dataset: %s\n', folders{f});

disp(head(data));

end

% Load the training data

trainingData = datasetTables.train;

testingData = datasetTables.test;

validData = datasetTables.valid;

%% Step 2: Verify Bounding Boxes and Labels

% Verify bounding box format

numRows = size(trainingData, 1); % Use size instead of height

disp("Number of rows in trainData: " + numRows);

for i = 1:numRows

flag_empty=true;

% if isempty(trainingData.boxes{i})

for jj=1:M

if ~isempty(trainingData.(classNames{jj}){i})

flag_empty=false;

break;

end

end

%if isempty(trainingData.boxes{i})

if flag_empty

error('Bounding boxes for image %s are missing.', trainingData.imageFilename(i));

end

% bbox = trainingData.boxes{i}{1}; % Extract bounding boxes for the current image

%if size(bbox, 2) ~= 4

% error('Bounding boxes for image %s are not in [x_min y_min width height] format.', trainData.imageFilename(i));

%end

end

% Verify labels are categorical

%for i = 1:numRows

% labels = trainingData.labels{i}{1}; % Extract labels for the current image

% if ~iscategorical(labels)

% error('Labels for image %s are not categorical.', trainingData.imageFilename(i));

% end

%end

% Verify boxes and labels are cell arrays

%if ~iscell(trainingData.boxes) || ~iscell(trainingData.labels)

% error('The boxes and labels columns must be cell arrays.');

%end

%% Step 3: Prepare R-CNN Training Pipeline

% Step 1: Extract Unique Classes

% Flatten the nested cell arrays

%flatLabels = cellfun(@(x) x{:}, trainingData.labels, 'UniformOutput', false);

% Concatenate all labels into a single categorical array

%labels = cat(1, flatLabels{:});

% Get unique classes

%uniqueClasses = unique(labels);

%numClasses = numel(uniqueClasses); % Number of unique classes

uniqueClasses=classNames;

numClasses = numel(uniqueClasses);

%disp(['Number of classes in trainData: ', num2str(numClasses)]);

% Add "background" class if not already present

if ~any(strcmp(uniqueClasses, "Background"))

uniqueClasses = [uniqueClasses; "Background"];

numClasses = numClasses + 1; % Increment numClasses to account for background

end

uniqueClasses

numClasses

% Load pre-trained ResNet-50

net = resnet50; % Use the pre-trained ResNet-50 model

% Create a layerGraph from the DAGNetwork

lgraph = layerGraph(net);

% Remove unnecessary layers (e.g., 'fc1000' and 'ClassificationLayer_fc1000')

lgraph = removeLayers(lgraph, {'fc1000', 'ClassificationLayer_fc1000', 'fc1000_softmax'});

% Add new layers for object detection

newFCLayer = fullyConnectedLayer(numClasses, 'Name', 'rcnnFC'); % numClasses includes background

lgraph = addLayers(lgraph, newFCLayer);

clear softmaxLayer;

% Add new softmax layer

softmaxLayerObj = softmaxLayer('Name','rcnnSoftmax'); % Use the built-in function

lgraph = addLayers(lgraph, softmaxLayerObj);

% Add new classification layer with explicit class names

newClassLayer = classificationLayer('Name', 'rcnnClassification', 'Classes', uniqueClasses);

lgraph = addLayers(lgraph, newClassLayer);

% Connect the layers

lgraph = connectLayers(lgraph, 'avg_pool', 'rcnnFC');

lgraph = connectLayers(lgraph, 'rcnnFC', 'rcnnSoftmax');

lgraph = connectLayers(lgraph, 'rcnnSoftmax', 'rcnnClassification');

% Verify the updated network

disp(lgraph.Layers);

analyzeNetwork(lgraph);

if(0) % the following code is removed....

%% Step 3: Visualize Bounding Boxes

% Visualize a few training samples

figure;

numSamples = 4; % Number of samples to visualize

numRows = size(trainingData, 1); % Get number of rows

idx = randperm(numRows, min(numSamples, numRows)); % Ensure it doesn't exceed available rows

for i = 1:numSamples

subplot(2, 2, i);

% Read the image

img = imread(trainingData.imageFilename(i));

% Get bounding boxes and labels for the current image

if isempty(trainingData.boxes{i}) || isempty(trainingData.labels{i})

error('Bounding boxes or labels for image %s are missing.', trainingData.imageFilename(i));

end

bbox = trainingData.boxes{i}{1}; % Extract bounding boxes

label = trainingData.labels{i}{1}; % Extract labels

% Display the image

imshow(img);

hold on;

% Draw bounding boxes and labels

for j = 1:size(bbox, 1)

rectangle('Position', bbox(j, :), 'EdgeColor', 'r', 'LineWidth', 2);

text(bbox(j, 1), bbox(j, 2) - 10, string(label(j)), ...

'Color', 'red', 'FontSize', 12, 'BackgroundColor', 'white', 'EdgeColor', 'black');

end

hold off;

end

end % of if (0)

%====================================================================

%=======================================================================

% Define training options

options = trainingOptions('sgdm', ...

'MiniBatchSize', 32, ...

'InitialLearnRate', 1e-4, ...

'MaxEpochs', 10, ...

'Verbose', true);

%% Step 5: Train the R-CNN Object Detector

% Train the R-CNN detector

detector = trainRCNNObjectDetector(trainingData, lgraph, options, ...

'NegativeOverlapRange', [0 0.3], ...

'PositiveOverlapRange', [0.6 1]);

disp('? Training Complete! Model successfully trained.');

%% Step 5: Test the Trained Detector

% Test the detector on a sample image

testImage = imread(testingData.imageFilename{1});

[bboxes, scores, labels] = detect(detector, testImage);

if isempty(bboxes)

warning('No objects detected in the test image.');

else

% Proceed with visualization

end

% Display the results

figure;

imshow(testImage);

hold on;

for i = 1:size(bboxes, 1)

rectangle('Position', bboxes(i, :), 'EdgeColor', 'r', 'LineWidth', 2);

text(bboxes(i, 1), bboxes(i, 2) - 10, string(labels(i)), ...

'Color', 'red', 'FontSize', 12, 'BackgroundColor', 'white', 'EdgeColor', 'black');

end

hold off;

Answer 1

[Cris LaPierre]

Not sure how your data is formatted, but typically this is done using the showShape function.

[Evangeline]

data is formatted in Darknet YOLO format of annotation files, and images dataset for training for Object detection. Multiple objects in an image

I have ran the above code for training, but step 3 is not executed, not sure how i can modify or i need to wait till the training is completed?

[Evangeline]

0 0.7113486842105263 0.881578947368421 0.019736842105263157 0.04111842105263158

0 0.7129934210526315 0.8453947368421053 0.019736842105263157 0.05263157894736842

0 0.6932565789473685 0.9037828947368421 0.02138157894736842 0.044407894736842105

0 0.6949013157894737 0.8733552631578947 0.02138157894736842 0.044407894736842105

0 0.696546052631579 0.8421052631578947 0.02138157894736842 0.044407894736842105

0 0.6735197368421053 0.9276315789473685 0.02138157894736842 0.046052631578947366

0 0.6751644736842105 0.897203947368421 0.023026315789473683 0.047697368421052634

0 0.6768092105263158 0.8651315789473685 0.023026315789473683 0.046052631578947366

0 0.6521381578947368 0.9539473684210527 0.024671052631578948 0.049342105263157895

0 0.6546052631578947 0.9226973684210527 0.024671052631578948 0.049342105263157895

0 0.6554276315789473 0.8898026315789473 0.024671052631578948 0.049342105263157895

0 0.6373355263157895 0.977796052631579 0.008223684210526315 0.044407894736842105

0 0.6274671052631579 0.946546052631579 0.03453947368421053 0.047697368421052634

0 0.6291118421052632 0.9161184210526315 0.03289473684210526 0.05263157894736842

0 0.7730263157894737 0.6381578947368421 0.014802631578947368 0.03289473684210526

0 0.774671052631579 0.609375 0.01644736842105263 0.03289473684210526

0 0.7919407894736842 0.649671052631579 0.027960526315789474 0.039473684210526314

0 0.7952302631578947 0.6200657894736842 0.027960526315789474 0.04111842105263158

0 0.796875 0.59375 0.027960526315789474 0.03453947368421053

0 0.8009868421052632 0.571546052631579 0.023026315789473683 0.023026315789473683

0 0.7549342105263158 0.6949013157894737 0.019736842105263157 0.03782894736842105

0 0.7574013157894737 0.665296052631579 0.019736842105263157 0.03782894736842105

0 0.759046052631579 0.6299342105263158 0.019736842105263157 0.047697368421052634

0 0.7393092105263158 0.7113486842105263 0.019736842105263157 0.039473684210526314

0 0.7417763157894737 0.680921052631579 0.019736842105263157 0.039473684210526314

0 0.7393092105263158 0.6480263157894737 0.027960526315789474 0.039473684210526314

0 0.740953947368421 0.6233552631578947 0.02138157894736842 0.024671052631578948

0 0.7384868421052632 0.3815789473684211 0.024671052631578948 0.044407894736842105

0 0.740953947368421 0.3412828947368421 0.02631578947368421 0.044407894736842105

0 0.7425986842105263 0.30180921052631576 0.02631578947368421 0.044407894736842105

0 0.740953947368421 0.2582236842105263 0.03453947368421053 0.046052631578947366

0 0.7179276315789473 0.38898026315789475 0.02631578947368421 0.047697368421052634

0 0.7203947368421053 0.3470394736842105 0.027960526315789474 0.047697368421052634

0 0.7220394736842105 0.3059210526315789 0.027960526315789474 0.047697368421052634

0 0.7195723684210527 0.26151315789473684 0.03618421052631579 0.049342105263157895

0 0.696546052631579 0.3980263157894737 0.027960526315789474 0.049342105263157895

0 0.6981907894736842 0.35444078947368424 0.027960526315789474 0.049342105263157895

0 0.6998355263157895 0.3100328947368421 0.027960526315789474 0.049342105263157895

0 0.6973684210526315 0.26480263157894735 0.03618421052631579 0.049342105263157895

0 0.6726973684210527 0.40625 0.027960526315789474 0.05263157894736842

0 0.6735197368421053 0.3618421052631579 0.029605263157894735 0.05098684210526316

0 0.6751644736842105 0.3157894736842105 0.03125 0.049342105263157895

0 0.6726973684210527 0.2680921052631579 0.039473684210526314 0.05263157894736842

0 0.6455592105263158 0.4161184210526316 0.03453947368421053 0.05263157894736842

0 0.6480263157894737 0.3684210526315789 0.03289473684210526 0.05263157894736842

0 0.6480263157894737 0.3223684210526316 0.03453947368421053 0.054276315789473686

0 0.6504934210526315 0.2722039473684211 0.03453947368421053 0.054276315789473686

1 0.3100328947368421 0.9802631578947368 0.07894736842105263 0.039473684210526314 for example, these are annotations for one such image

[Cris LaPierre]

Consider saving your variables to a mat file and attaching that to your post using the paperclip icon.

It would be helpful to have at least one image and its corresponding bounding boxes and labels.

[Cris LaPierre]

I ended up finding a Darknet example data set to use here.

Of note

• Darknet annotations are stored with the format class_id center_x center_y width height
• Darknet annotations are normalized to be between [0,1]. You need to convert to pixels.
• showShape expects a 5th column for yaw when defining rectangle location using xctr and yctr

Here's how I would do it using the example I found online.

unzip('air4.zip')

% Load image
img = imread('air4.jpg');

% Load classnames
classNames = readlines('obj.names');

% Load annotations
obj = readmatrix('air4.txt');

% extract labels
labelID = obj(:,1)+1;

% Scale position coordinates using image dimensions
bboxes = obj(:,2:5).*[size(img,[2 1 2 1])];
% Add column for yaw so showShape knows position is defined as (xctr,yctr)
bboxes(:,5)=0;

% Display labeled image
imshow(img)
showShape("rectangle",bboxes,Label=classNames(labelID))

●