Shifting Weights: Adapting Object Detectors from Image to Video