Our dataset includes three parts: MNIST-rot,  MNIST-scale, and MNIST-rand.
MNIST-rot is generated by randomly rotating
each sample in the MNIST testing dataset in $[0,2\pi]$.
We generated MNIST-scale by randomly scaling
the ratio of the area occupied by the symbol over that of the entire image by a factor
in $[0.5,1]$, and generated MNIST-rand by scaling and
rotating images in MNIST testing dataset simultaneously.