How to increase precision of image processing functions?

I have a 4K video which I stabilize with OpenCV and get a perfect result. Since I need to work with some stabilized frames in Matlab, I saved every transformation Matrix (10 significant digits) for every frame calculated by OpenCV in a CSV file. Then I read the CSV file in Matlab and apply the transformation to the frame I want. Minimum Code example I use:

affineTrafo = affine2d(transpose(transformMat));
[transformedFrame, transformedRef] = imwarp(frame, affineTrafo);
[xIntrinsic, yIntrinsic] = worldToIntrinsic(transformedRef, 0, 0);
padSize = ceil(max(abs(xIntrinsic), abs(yIntrinsic))) + 1;
paddedFrame = padarray(transformedFrame, padSize, 'both');
croppedFrame = imcrop(paddedFrame, [xInstrinsic + padSize, yIntrinsic + padSize, 4095, 2159]);
result = croppedFrame;

The problem is now, while my result generated with OpenCV produces a perfect stabilized video, Matlab produces with the exact same data some kind of 'swimming'. The reference pixels aren't holding their position but move 1-2 pixels all the time, thus making the result useless.

I thought maybe the non integer numbers in the cropping are the problem so I changed my code to this:

flooredX = floor(xIntrinsic);
flooredY = floor(yIntrinsic);
fracX = xIntrinsic - flooredX;
fracY = yIntrinsic - flooredY;
fracTranslation = affine2d([1, 0 , 0; 0, 1, 0; fracX, fracY, 1]);
translatedFrame = imwarp(transformedFrame, fracTranslation);
padSize = 100;
paddedFrame = padarray(translatedFrame, padSize, 'both');
croppedFrame = imcrop(paddedFrame, [flooredX + padSize, flooredY + padSize, 4095, 2159]);
result = croppedFrame;

but the results are exactly the same.

It seems that this is an precision problem, but I can't see where Matlab could make numeric errors in this process. Does someone has a clue why the result differs from OpenCV?