A complementary dual model for weakly supervised salient object detection

Liyuan Chen, Dawei Zhang, Xiao Wang, Chang-Dong Wang, Shan Jin, Zhonglong Zheng

Abstract: Highlights•CRANet integrates complementary information from foreground and background models.•A pseudo-label aggregation strategy enhances the supervision of scribble annotations.•A flipping consistency method improves prediction consistency.•A flipped guided loss expands the effective training set.•The performance of CRANet significantly outperform existing methods.