Creating Image Datastore as a struct for codegen throwing an error

Hi,

I am trying to create an ImageDatastore object to initialize a bag of features and then run codegen on it. Documentation says I can create it as a struct but it still throws an error.

images = cell(numImages, 1);

labels = cell(numImages, 1);

for k = 0:numImages-1

filename = fullfile(imageFolder, sprintf('%04d.jpg', int32(k)));

images{k+1} = imread(filename);

labels{k+1} = sprintf('%04d.jpg', int32(k));

end

imds.Images = images;

imds.Labels = categorical(labels);

bag = bagOfFeatures(imds, 'CustomExtractor',CustomFeatureExtractor, ...

bowParams{:});

This is the error I get:

• It no longer gives me an error on codegen.
• It doesn't work with regular execution.

Error using bagOfFeatures/parseInputs (line 1005)

The value of 'imds' is invalid. Expected imds to be one of these types:

matlab.io.datastore.ImageDatastore

Instead its type was struct.

Error in

bagOfFeatures (line 119)

[imgSets, params] = bagOfFeatures.parseInputs(varargin{:});

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Error in

main_vSLAM_VisualHull_RaspPi (line 283)

bag = bagOfFeatures(imds, 'CustomExtractor',CustomFeatureExtractor, ...

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Am I doing something wrong here?

Answer 1

[Sahithi Kanumarlapudi]

Hi Sankar,

You're not doing anything wrong — this is expected behavior. The struct-based input is only supported in the code generation path, not during regular MATLAB execution.

When you run bagOfFeatures in MATLAB directly, it requires an actual ImageDatastore object. The codegen implementation accepts a struct with:

- Images — cell array of images (numeric arrays from imread)

- Labels (optional) — categorical vector

So your struct format is correct for codegen. For regular MATLAB execution (e.g., testing/prototyping), use an actual ImageDatastore:

imds = imageDatastore(imageFolder, 'FileExtensions', '.jpg');

bag = bagOfFeatures(imds, 'CustomExtractor', CustomFeatureExtractor, bowParams{:});

If you need the same code to work in both paths, you can use coder.target to branch:

if coder.target('MATLAB')

imds = imageDatastore(imageFolder, 'FileExtensions', '.jpg');

else

imds.Images = images;

imds.Labels = categorical(labels);

end

bag = bagOfFeatures(imds, 'CustomExtractor', CustomFeatureExtractor, bowParams{:});

This is the designed workflow — the struct is a codegen-only substitute since ImageDatastore objects are not supported for

code generation