Abstract: Building an effective object detector usually depends on large well-annotated training samples. While annotating such dataset is extremely laborious and costly, where box-level supervision ...