Learning with Category-Equivariant Representations for Human Activity Recognition