Two-Stage Training for Improved Classification of Poorly Localized Object Images