The regression head det-reg is trained only one seedbounding box annotations B0, too, so that it is not affectedby the inaccurate box coordinates in Bt.Filtering background proposal loss. Given that falsenegative is extremely hard to eliminate in mined boundingboxes, the losses of “background” proposals in det-cls-a arenot used in loss to remove the effect of false negatives.Specifically, if an image i is from mined box annotationset Bt, then we mask the losses from the proposals thatbelong to “background” category (typical implementationuses index 0 for background); if the image is from seed box