The MultiMNIST training and test dataset is generated by overlaying a digit on top of another digit from the same set (training or test) but different class.

Data Collection

Each digit is shifted up to 4 pixels in each direction resulting in a 36¡Á36 image. Considering a digit in a 28¡Á28 image is bounded in a 20¡Á20 box, two digits bounding boxes on average have 80% overlap. For each digit in the MNIST dataset we generate 1K MultiMNIST examples. So the training set size is 60M and the test set size is 10M.