Question about MATLAB Mask RCNN code from github.

Hello, I am trying to test the code, 'Mask RCNN with MATLAB' on the external website, 'https://github.com/matlab-deep-learning/mask-rcnn'.

I have a question on the function, [rpnTrainingSamples.m] , 41~50 line.

The code is :

=====================================================================================

% Sub-sample negative samples to avoid using too much memory.

numPos = sum(positiveIndex);

negIdx = find(negativeIndex);

numNeg = numel(negIdx);

nidx = params.RandomSelector.randperm(numNeg, min(numNeg, 5000));

% Pack data as int32 to save memory.

regionProposals = int32([regionProposals(positiveIndex, :); regionProposals(nidx, :)]);

anchorIDs = {int32([anchorIDs(positiveIndex) anchorIDs(nidx)])};

anchorIndices = {int32([anchorIndices(positiveIndex,:); anchorIndices(nidx,:)])};

=====================================================================================

The above packed data will be used for training samples of RPN.

However, for the negative samples of regionProposals (and anchorIDs, etc..), those with indices 'nidx' are packed.

As I know of, the variable 'nidx' is just a randomly shuffled numbers in the range of [1, numNeg].

So, I am afraid that 'regionProposals(nidx, :)' may not represent the negative samples but all the samples (positive, negative, ignored).

Is it right?

Thank you!

Answer 1

[Prabhan Purwar]

Hi,

Thanks for reporting the issue. As a workaround kindly consider replacing nidx with numNeg(nidx) in the above code. I have bought the issue to the notice of our developer and will try to fix it in future releases.

Thanks