Moving MNIST

Moving MNIST is a benchmark data set for video recognition. There are 10,000 samples including 8,000 for training and 2,000 for test. Each sample consists of 20 sequential gray scale images of 64 x 64 patch, where two digits move in the sequences.