Semantic-Aware Superpixel for Weakly Supervised Semantic Segmentation

Weakly-supervised semantic segmentation aims to train a network using weak labels. Among labels, image-level label has been the most popular choice due its simplicity. However, since labels lack accurate object region information, additional modules such as saliency detector have exploited in weakly supervised segmentation, which requires pixel-level for training. In this paper, we explore self-supervised vision transformer mitigate heavy efforts on generation of annotations. By exploiting features obtained from transformer, our superpixel discovery method finds out semantic-aware superpixels based feature similarity an unsupervised manner. Once obtain superpixels, superpixel-guided seeded growing method. Despite simplicity, approach achieves competitive result with state-of-the-arts PASCAL VOC 2012 and MS-COCO 2014 datasets segmentation. Our code is available at https://github.com/st17kim/semantic-aware-superpixel.