A Study On Robustness And Semantic Understanding Of Visual Models