A Weakly Supervised Learning Framework for Salient Object Detection via Hybrid Labels

Fully-supervised salient object detection (SOD) methods have made great progress, but such often rely on a large number of pixel-level annotations, which are time-consuming and labour-intensive. In this paper, we focus new weakly-supervised SOD task under hybrid labels, where the supervision labels include coarse generated by traditional unsupervised method small real labels. To address issues label noise quantity imbalance in task, design pipeline framework with three sophisticated training strategies. terms model framework, decouple into refinement sub-task sub-task, cooperate each other train alternately. Specifically, R-Net is designed as two-stream encoder-decoder equipped Blender Guidance Aggregation Mechanisms (BGA), aiming to rectify for more reliable pseudo-labels, while S-Net replaceable network supervised pseudo current R-Net. Note that, only need use trained testing. Moreover, order guarantee effectiveness efficiency training, strategies, including alternate iteration mechanism, group-wise incremental credibility verification mechanism. Experiments five benchmarks show that our achieves competitive performance against weakly-supervised/unsupervised both qualitatively quantitatively. The code results can be found from link https://rmcong.github.io/proj_Hybrid-Label-SOD.html .