Human motion convolutional autoencoders using different
rotation representations