Towards Cross-View Consistency in Semantic Segmentation While Varying View Direction