Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation

• DOI: 10.1609/aaai.v38i3.27980
• Publication Date: 2024-03-25T09:18:01Z
• Authors: Jingxuan He; Lechao Cheng; Chaowei Fang; Zunlei Feng; Tingting Mu; Mingli Song
• Keywords: Feature (linguistics); Semantic feature

Compared to conventional semantic segmentation with pixel-level supervision, weakly supervised (WSSS) image-level labels poses the challenge that it commonly focuses on most discriminative regions, resulting in a disparity between and fully supervision scenarios. A typical manifestation is diminished precision object boundaries, leading deteriorated accuracy of WSSS. To alleviate this issue, we propose adaptively partition image content into certain regions (e.g., confident foreground background) uncertain boundaries misclassified categories) for separate processing. For cues, an adaptive masking strategy seek recover local information self-distilled knowledge. We further assume should be robust enough preserve global semantics, introduce complementary self-distillation method constrains consistency augmented view same class labels. Extensive experiments conducted PASCAL VOC 2012 MS COCO 2014 demonstrate our proposed single-stage approach WSSS not only outperforms state-of-the-art counterparts but also surpasses multi-stage methods trade complexity accuracy.