How to create a box label datastore?

Hello, I've been trying to create a box label datastore using the blds function.

What I understood, is that I need to create a table with the first column containing a type 'double' array, coordinates of bounding boxes, and a string variable with its class.

Now, how can I assign a 1x4 'double' vector into 1 single cell? In this 'fake' case, coordinates of bb are all the same, so is the class. I would need something like:

[0, 0, 300, 300] 'stopSignal'

[0, 0, 300, 300] 'stopSignal'

bbox = [0 0 300 300]; %images: 4170
maxSamples = 4170;
varTypes = {'double', 'string'};
size = [4170 2];
T = table('size',size, 'VariableTypes', varTypes);
for i = 1:maxSamples
T (i,1) = {bbox};
end

The error occurs because bbox is a 1x4 double vector, while MATLAB expects only a 1x1 variable to be stored in one single cell.

What am I missing?

Answer 1

[Abolfazl Chaman Motlagh]

you can easily create and edit a cell. then convert it to table.

here is an example:

boxes = cell(10,2); %number of images x 2=(coordinates of box , labels)
% fill boxes :
for i=1:10
n = randi(3); % number of box in i-th image, it maybe diffrenent so i consider it
boxes{i,1} = rand(n,4); % nx4 each row coordinate of a box
boxes{i,2} = string(randi(2,n,1)); % here i create n label for every image between 2 possible labels
end

% Convert to table
boxes = cell2table(boxes,'VariableNames',{'Boxes','Labels'}); 

blds = boxLabelDatastore(boxes)

blds = 
  boxLabelDatastore with properties:

    LabelData: 
               10x2 cell array
               
               {[0.1264 0.5256 0.1346 0.2392]}    {[1            ]}
               {[0.3028 0.0742 0.1592 0.0875]}    {[2            ]}
               {3×4 double                   }    {3×1 categorical}
               
                ... and 7 more rows
    ReadSize: 1

Remember if this blds is gonna use for a deep learning applications, values of each box should be checked. 0 is invalid for deep learning. and sum of 3-th element and first one shouldn't be more than image width, and sum of 4th and 2th element shouldn't be more than image height.

[Lorenzo Tonelli]

Thank you for your answer! But sadly this didn't fix my problem: if I run your code, everything is perfect. But in my case, instead of using:

boxes {i,1} = rand(n,4);

I'm using:

boxes {i,1} = [2 2 298 298];

and this gives the following error:

Error using

boxLabelDatastore>iAssertValidBBoxFormat (line

861)

The size of bounding box data must be M-by-4,

M-by-5, or M-by-9, where M is the number of boxes

in each table element. The column in the training

data table that contains the bounding boxes must

be a cell array.

I don't understand. Using rand gives an array. My imput is an array. What's the difference?

My table:

Table created by your code:

There's a pretty clear difference. I don't understand why!

[Abolfazl Chaman Motlagh]

Oh yes. when all boxes are 1x4, the cell2table automatically change the table first column to double not cell.

so here's the solution:

create boxes as cell, then use table function.

for i=1:10
boxes(i,1) = {[2 2 298 298]};
end
labels = num2cell(string(randi(2,10,1)));

boxes = table(boxes,labels);

blds = boxLabelDatastore(boxes)

blds = 
  boxLabelDatastore with properties:

    LabelData: 
               10x2 cell array
               
               {[2 2 298 298]}    {[2]}
               {[2 2 298 298]}    {[2]}
               {[2 2 298 298]}    {[1]}
               
                ... and 7 more rows
    ReadSize: 1

[Lorenzo Tonelli]

Thank you! Works like a charm.