Referring perception models need to be robust against various perturbations for reliable real-world applications.
