Semantic Depth Matters: Explaining Errors of Deep Vision Networks through Perceived Class Similarities