Bridging Visual Representations for Object Detection