Avoiding memory reallocation in Matlab

Question

I am converting an implementation of an image segmentation method from C++ to Matlab.

It is an iterative method that that make calls to two functions: em() and mpm(). Those two functions uses two reasonably big auxiliary matrices, so to avoid reallocating memory at every call in my C++ implementation, I have prealocated the memory for the matrices and reseted its values when it was necessary.

How can I avoid memory reallocation in the Matlab implementation? As far as I know, pre-allocating the matrices won't solve my problem because when a function tries to change the values of an argument, Matlab automatically creates a copy of it.

Is it possible to declare a variable that is shared by all functions inside a M file?

Edit:

Here's an outline of my C++ code. The two matrices that I preallocate are chanceMatrix and labelAssignedVector:

GSegmentedImage * GEmMpmSegmentator::segmentImage(GImage * image, int labelRange, int steps, unsigned char minGrayLevel){
GSegmentedImage * segmentedImage = new GSegmentedImage(image->getHeight(), image->getWidth());
GParameterVector * paramVector = new GParameterVector(labelRange);

randomizeLabelField(segmentedImage, labelRange);
setOutOfImageLabel(segmentedImage, image, minGrayLevel);
initializeParameterVector(image, paramVector);

// labelAssignedVector stores how many times a pixel received a label 'k'.
// row = image row, column = image column, page = label.
G3DMatrix<unsigned char> * labelAssignedVector = new G3DMatrix<unsigned char>(image->getHeight(), image->getWidth(), labelRange);

// Chance matrix is used by the mpm method.
// chanceMatrix: row = label, column = gray level, page = number of different neighbors
G3DMatrix<double> * chanceMatrix = new G3DMatrix<double>(paramVector->getLabelRange(), 256, 9);

for (int i = 0; i < steps; i++){
    labelAssignedVector->reset();
    if (difPenalty < difMax){
        difPenalty += difInc;
    }

    mpm(paramVector, labelAssignedVector, segmentedImage, image, minGrayLevel, chanceMatrix);
    em(paramVector, labelAssignedVector, image);
}

delete paramVector;
return segmentedImage;
}

Answer 1

While the global command can declare a shared variable, this usually isn't a good idea.

One possibility would be to create a handle class that holds some stuff. Perhaps something like:

ValueHandle.m:

classdef ValueHandle < handle
    properties
        value
    end
    methods
        function self = ValueHandle(value)
            self.value = value;
        end
    end
end

Now you can create ValueHandle objects and access their value property. The handle itself can be passed between functions like a reference or pointer in another language. This might help somewhat. This probably requires a recentish version of MATLAB.

If this doesn't help as much as you'd like, have you considered using MATLAB's mex extension system to bridge your existing C++ code with MATLAB?

Answer 2

Have a look at the following stackoverflow post.