Exploring Why Object Recognition Performance Degrades Across Income Levels and Geographies with Factor Annotations