Codegen for Bundle Adjustment

I am facing two issues when using codegen on my code. I have attached images of my code for reference.

1) The error report recognises cameraPoses as a struct even though it is a 2x2 table when I run the code. When I take the values in cameraPoses and make camPoses (as shown in the image), that is recognised as a table.

2) Once I use camPoses in Bundle Adjustment, I get a codegen error that says 'This function does not support one-dimensional indexing. I am not sure what that means because my variable dimensions match the documentation. I have also attached images of my variables and their dimensions.

Answer 1

[Matt]

Hi @Sankar,

The issue is not your runtime dimensions. Instead, tt is a codegen representation/type issue.

In MATLAB R2026a, the codegen implementation of imageviewset/poses returns a struct:

struct('ViewId', viewsId, 'AbsolutePose', absPoses)

So the report calling "cameraPoses" a struct is expected, even though normal MATLAB execution shows a 2x2 table.

cameraPoses = poses(vSetKeyFrames);
 
viewIds = uint32(cameraPoses.ViewId(:));
numViews = numel(viewIds);

absolutePose = cell(numViews, 1);
for k = 1:numViews
    absolutePose{k} = cameraPoses.AbsolutePose(k);
end

camPoses = table(viewIds, absolutePose, ...
    'VariableNames', {'ViewId', 'AbsolutePose'});

[refinedPoints, refinedAbsPoses] = bundleAdjustment( ...
    xyzWorldPoints, tracks, camPoses, intrinsics, ...
    'FixedViewIDs', uint32(1), BAParams.full{:});

Additionally, I recommend avoiding "1:height(cameraPoses)" for codegen here. If "cameraPoses" is the codegen struct, "height(cameraPoses)" is not the number of poses. Use "numel(cameraPoses.ViewId)" and preserve the actual ViewId values so they still match "tracks.ViewIds".

“One-dimensional indexing” means codegen is seeing an object array that only supports linear indexing like "pose(k)", but something in the generated validation path is trying table/row-style indexing. Making "AbsolutePose" a cell column avoids that path failing.

Hope this resolves the issue.

Regards,

Matt.

[Sankar]

That solved the issue! Thank you so much.