Weakly Supervised Salient Object Detection via Hybrid Pseudo Labels
Yixin Liu, Fengqi Liu, Miao Zhang · 2024
Weakly supervised salient object detection (WSOD) can achieve up to 90% of the performance of fully supervised salient object detection by using low-cost annotations. Most existing WSOD methods focus on utilizing single annotations and enhancing saliency networks. Although these methods have achieved outstanding performance, the single pseudo label tends to exhibit one-sided characteristics, thereby impacting the saliency networks. In this work, we first propose a weakly supervised salient object detection method that combines image-level and scribble annotations, effectively leveraging the strengths of both types of annotation. Specifically, we design a dual-branch network framework, with one branch generating image-level pseudo labels and the other generating scribble pseudo labels. Also, we design a cross fusion module (CFM) which significantly improves the accuracy and robustness of saliency detection by efficiently integrating feature information from different levels and different sources of pseudo labels. Extensive experiments conducted on five widely-adopted salient object detection datasets provide evidence of the effectiveness and superiority of the method proposed in this study. Our method offers a new perspective in the field of salient object detection, proving the potential of combining different types of annotation information and significantly contributing to the performance enhancement of weakly supervised learning methods.