Double-Head RCNN: Rethinking Classification and Localization for Object Detection

A Double-Head method is proposed, which has a fully connected head focusing on classification and a convolution head to pay more attention to bounding box regression.