Detecting Systematic Weaknesses in Vision Models along Predefined Human-Understandable Dimensions