Towards Verifying the Geometric Robustness of Large-scale Neural Networks