Error using minibatchque when training pointPilla​rsObjectDe​tector.

I am trying to train a pointPillarsObjectDetector on a small set of PCD files and corresponding ground truth labels from the Lidar Labeler App. I largely followed the "Lidar 3-D Object Detection Using PointPillars Deep Learning" example and made only a few changes. When I go to train the detector, I get a minibatchqueue error. I supplied the error message and my code below (I am not allowed to supply the data; it is just a folder of PCD files and exported ground truth data from the Lidar Labeler App). Does anyone have any idea why I might be getting the error? Thanks.

*************************************************************************

Processing data in minibatchqueue....

Error using minibatchqueue (line 188)

Unable to apply function specified by 'MiniBatchFcn' value.

Error in trainPointPillarsObjectDetector (line 52)

mbq = minibatchqueue(...

Error in detector (line 91)

[objDetector,info] = trainPointPillarsObjectDetector( ...

Caused by:

Error using sub2ind (line 71)

Out of range subscript.

%% Create a file datastore

% Path to PCD files
pcdPath = '/Users/admin/Desktop/pyCode/lidarExploration/pcdEx';
% Create a file datastore
lds = fileDatastore(pcdPath, 'ReadFcn', @(x) pcread(x), "FileExtensions", '.pcd');



%% Create a box label datastore

% Load ground truth object
gt = load('labelsRun2.mat');
% Extract the label data
labels = timetable2table(gt.gTruth.LabelData);
% Take the bounding boxes
bboxes = labels(:, 2);
% Make a box label datastore
bds = boxLabelDatastore(bboxes);



%% Combine the datastores

% Combine the lidar and box datastores
cds = combine(lds, bds);



%% Estiamte anchor boxes

% Create anchor boxes for each class
anchorBoxes = calculateAnchorsPointPillars(bboxes);
% Record the class names
classNames = bboxes.Properties.VariableNames;



%% Define the region of interest

xMin = -30;     % Minimum value along X-axis.
yMin = -50;  % Minimum value along Y-axis.
zMin = -10;    % Minimum value along Z-axis.
xMax = 62.16;   % Maximum value along X-axis.
yMax = 55;   % Maximum value along Y-axis.
zMax = 0;     % Maximum value along Z-axis.
xStep = 0.16;   % Resolution along X-axis.
yStep = 0.16;   % Resolution along Y-axis.

% Define point cloud parameters.
pointCloudRange = [xMin xMax yMin yMax zMin zMax];
voxelSize = [xStep yStep];

% Calculate dimensions for the pseudo-image
Xn = round((xMax - xMin) / xStep);
Yn = round((yMax - yMin) / yStep);

% Helper function to check pseudo-image dimensions
function helperValidatePointCloudRange(Xn, Yn)
    if (mod(Xn, 4) ~= 0) || (mod(Yn, 4) ~= 0)
        error("Define PointCloudRange and VoxelSize such that Xn and Yn are multiple of 4")
    end
end

% Validate pseudo-image dimensions
helperValidatePointCloudRange(Xn, Yn);



%% Define the object detector

% Define the object detector
objDetector = pointPillarsObjectDetector( ...
    pointCloudRange, ...
    classNames, ...
    anchorBoxes, ...
    VoxelSize = voxelSize ...
    );



%% Train the object detector

% Training options
options = trainingOptions( ...
    'adam', ...
    'MaxEpochs', 10 ...
    );

% Train the detector
[objDetector,info] = trainPointPillarsObjectDetector( ...
    cds, ...
    objDetector, ...
    options ...
    );

Answer 1

[Gayathri]

Hi @Clayton,

I was able to reproduce the error in MATLAB R2024a version. I had used the same input data present in "Lidar 3-D Object Detection Using PointPillars Deep Learning" example. You can refer to this documentation link for the example:

https://www.mathworks.com/help/deeplearning/ug/lidar-object-detection-using-pointpillars-deep-learning.html

I tried changing the ROI values to the values given in the "Lidar 3-D Object Detection Using PointPillars Deep Learning" example with which the error was resolved. I am referring to the below mentioned parameters.

xMin = 0.0;     % Minimum value along X-axis.
yMin = -39.68;  % Minimum value along Y-axis.
zMin = -5.0;    % Minimum value along Z-axis.
xMax = 69.12;   % Maximum value along X-axis.
yMax = 39.68;   % Maximum value along Y-axis.
zMax = 5.0;     % Maximum value along Z-axis.
xStep = 0.16;   % Resolution along X-axis.
yStep = 0.16;    % Resolution along Y-axis.

The range mentioned in your code might not be aligning with your actual data, leading to issues, when trying to index into arrays or matrices. I suggest you try adjusting the values to align with your data.

Hope you find this information helpful.

[Clayton]

Thanks @Gayathri! That seemed to solve the issue. My point clouds were already cropped to a certain size, so I speficied those same limits.

xMin = -30;     % Minimum value along X-axis.
yMin = -50;  % Minimum value along Y-axis.
zMin = -6;    % Minimum value along Z-axis.
xMax = 62;   % Maximum value along X-axis.
yMax = 55;   % Maximum value along Y-axis.
zMax = 0;     % Maximum value along Z-axis.

Then, I changed the resolutions to match the psuedo-image dimensions (432 x 496) from the "Lidar 3-D Object Detection Using PointPillars Deep Learning" example.

xStep = (xMax - xMin) / 432;   % Resolution along X-axis.
yStep = (yMax - yMin) / 496;   % Resolution along Y-axis.

I also changed ResetInputNormalization to match the example.

% Training options
options = trainingOptions( ...
    'adam', ...
    'MaxEpochs', 10, ...
    'ResetInputNormalization', false ...
    );