Learning with Category-Equivariant Architectures for Human Activity Recognition