Spherical MNIST

Spherical MNIST is constructed from the MNIST dataset by back projecting the digits into equirectangular projection with a resolution of 160x80. The digit labels are used to train the source CNN (recognition model), but they are not used to train the KTN.