The four dashed lines serve as baselines. Blue lines denote co-training without BBAug and green lines are co-training with BBAug. The symbol squares are results on clean data and the triangles are results on noisy data. After adding BBAug, the mAP on the clean dataset goes down in general while the mAP on the noisy dataset goes up. Taking all this into account, the point with the threshold of prediction confidence at 0.5 and added with BBAug can meet the expectations of both clean and noisy datasets