BBAM: Bounding Box Attribution Map for Weakly Supervised Semantic and Instance Segmentation

Weakly supervised segmentation methods using bounding box annotations focus\non obtaining a pixel-level mask from each box containing an object. Existing\nmethods typically depend on a class-agnostic mask generator, which operates on\nthe low-level information intrinsic to an image. In this work, we utilize\nhigher-level information from the behavior of a trained object detector, by\nseeking the smallest areas of the image from which the object detector produces\nalmost the same result as it does from the whole image. These areas constitute\na bounding-box attribution map (BBAM), which identifies the target object in\nits bounding box and thus serves as pseudo ground-truth for weakly supervised\nsemantic and instance segmentation. This approach significantly outperforms\nrecent comparable techniques on both the PASCAL VOC and MS COCO benchmarks in\nweakly supervised semantic and instance segmentation. In addition, we provide a\ndetailed analysis of our method, offering deeper insight into the behavior of\nthe BBAM.\n