Hello there. I'm not exactly sure about the idx3-ubyte format but I guess that's only a matter of converting, lets say, the information stored as doubles to this format.
What I think that should be of any help to you is to first check the internal way that the original mnist dataset organizes the images. For example: I am currently working with mnist dataset used in PCANet. Internally the images are structured as vectors. So, for the train part of this dataset, the fisrt row of the matrix that holds all the training images is image1, the second row is image2, and so on, with each label appended at the end of the vectorized image.
I had to build my own dataset as well, so I had to preprocess then in order to have then with the same size, and then build a matrix with each vectorized image corresponding to a row in that matrix. Check the image below:
You can see that mnist_basic.mat has in the test dataset 50000 images, each one corresponding to a vector with 785 positions: the first 784 correspond to the image itself (28x28) and the last one is its label.
Heres a simple code to do that:
n_samples = 3;
data(1).im = imread('image1.png');
data(2).im = imread('image2.png');
data(3).im = imread('image3.png');
[w, d] = size(data(1).im);
data_train = zeros(n_samples, w*d);
for i = 1: n_samples
image = reshape(data(i).im, [1 w*d]);
im_aux_label = [image,lable]
data_train(i, :) = im_aux_label;
end
save('my_dataset.mat','data_train');
Hope it helps you somehow.
