Inter-image Token Relation Learning for weakly supervised semantic segmentation

Jingfeng Tang

doi:10.1016/j.jvcir.2025.104576

Inter-image Token Relation Learning for weakly supervised semantic segmentation

2025-09-22

0

PRE

AI

DOI：10.1016/j.jvcir.2025.104576

Original

Original request for help

Save

Abstract

En 中文

• The Inter-image Class Token Contrast scheme contrasts semantic representations across images and constrains inter-image attention, activating more semantic regions in naive attention maps from the c2p step. • The Inter-image Patch Token Align aligns low-level features across images using single-class labels, enhancing patch token mutual information and strengthening dependencies for p2p optimization step. • The proposed method enhances inter-image token relation learning, achieving competitive mIoU on PASCAL VOC 2012 and MS COCO 2014 in WSSS task.

Inter-image Token Relation Learning for weakly supervised semantic segmentation

Abstract

Journal

Organization

Cited Papers