Structure-Consistent Weakly Supervised Salient Object Detection with Local Saliency Coherence

Sparse labels have been attracting much attention in recent years. However, the performance gap between weakly supervised and fully salient object detection methods is huge, most previous works adopt complex training with many bells whistles. In this work, we propose a one-round end-to-end approach for via scribble annotations without pre/post-processing operations or extra supervision data. Since fail to offer detailed regions, local coherence loss propagate unlabeled regions based on image features pixel distance, so as predict integral complete structures. We design saliency structure consistency self-consistent mechanism ensure consistent maps are predicted different scales of same input, which could be viewed regularization technique enhance model generalization ability. Additionally, an aggregation module (AGGM) better integrate high-level features, low-level global context information decoder aggregate various information. Extensive experiments show that our method achieves new state-of-the-art six benchmarks (e.g. ECSSD dataset: F? = 0.8995, E? 0.9079 MAE 0.0489), average gain 4.60% F-measure, 2.05% E-measure 1.88% over best performing task. Source code available at http://github.com/siyueyu/SCWSSOD.