An example with the setting of corruption = fog, severity = 1. Our method surpasses Faster-RCNN by detecting more objects including cyclists and pedestrians. Also, BBAug greatly boosts the robustness compared with the counterpart without BBAug: (a) Ground-truth annotation, (b) result of Faster-RCNN with supervised pretrained model, (c) result of Faster-RCNN with self-supervised pretrained MoCo, (d) result of Co-training without BBAug, (e) result of Co-training with BBAug