Cyclic self-guidance for weakly-supervised instance segmentation

Abstract

We propose a novel approach for the weakly-supervised instance segmentation using only image-level labels. We aim to properly reflect instance-level information without off-the-shelf proposal techniques that most of the existing works depend on. First, we exploit CAMs to generate initial pseudo labels representing each instance as a center point and 2D offset vectors. To generate more reliable labels from CAMs, we present two modules that control the discriminative region and produce dense and sparse CAMs, respectively. Second, we propose a self-learning strategy named a \textit{cyclic self-guidance} that allows the network to produce refined pseudo labels and employ them for training in a self-supervised and cyclic learning manner. Extensive experiments on PASCAL VOC 2012 show the effectiveness of our approach in instance segmentation only using image-level labels. On the instance segmentation benchmark, our new attempt in this field achieves 41.5% $mAP_{50}$ surpassing existing works without off-the-shelf proposal techniques. Also, we achieve a comparable performance of 67.1% mIoU on the semantic segmentation benchmark.