.png){kind=icon}
.png){kind=icon}
Return
Language-guided modulation-update for semi-supervised semantic segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.113505.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We propose LMU, a novel semi-supervised semantic segmentation framework that incorporates language-guided conditional feature updates to better exploit the relationship between labeled and unlabeled data. • We design a semantic-driven adaptive feature condition update strategy that enforces class-consistent feature interactions, alleviating the ambiguity introduced by indiscriminate conditional updates. • We propose a Category Semantic Understanding Module, which exploits the refined similarity matrix to modulate unlabeled features, thereby ensuring a more reliable and semantically consistent conditional updating process. • We conduct extensive experiments to rigorously validate the effectiveness of LMU. The results consistently demonstrate significant improvements over stateof- the-art methods on three widely recognized benchmarks: PASCAL VOC 2012, Cityscapes, and COCO.
Keywords:
LMU
semi-supervised semantic segmentation
language-guided feature update
adaptive feature condition
category semantic understanding