Evaluating Context for Deep Object Detectors