Abstract: Well-annotated training samples show necessity in achieving high performance of object detection, but collection of massive samples is extremely laborious and costly. Recently, cut-paste ...