Lee Jungbeom ; Yi Jihun ; Shin Chaehun ; Yoon Sungroh

Abstract

Weakly supervised segmentation methods using bounding box annotations focuson obtaining a pixel-level mask from each box containing an object. Existingmethods typically depend on a class-agnostic mask generator, which operates onthe low-level information intrinsic to an image. In this work, we utilizehigher-level information from the behavior of a trained object detector, byseeking the smallest areas of the image from which the object detector producesalmost the same result as it does from the whole image. These areas constitutea bounding-box attribution map (BBAM), which identifies the target object inits bounding box and thus serves as pseudo ground-truth for weakly supervisedsemantic and instance segmentation. This approach significantly outperformsrecent comparable techniques on both the PASCAL VOC and MS COCO benchmarks inweakly supervised semantic and instance segmentation. In addition, we provide adetailed analysis of our method, offering deeper insight into the behavior ofthe BBAM.