This project provides the code and results for "ORSI Salient Object Detection via Progressive Semantic Flow and Uncertainty-aware Refinement", IEEE TGRS, vol. 62, pp. 5608013–5608025, 2024. If you have any questions, please feel free to contact us："qyq@zjut.edu.cn" or "zjw@zjut.edu.cn"
In this paper, we propose SFANet, a new model for salient object detection in remote sensing images. SFANet mainly comprises four new modules: Semantic Extraction Module (SEM), Interscale Fusion Module (IFM), Deep Semantic Graph-Inference Module (DSGM), and Uncertainty-aware Refinement Module (URM), which address various challenges in RSI-SOD. Specifically, SEM extracts semantic flow knowledge from low-level features, IFM refines and fuses low-level semantic flow information, and DSGM exploits deep semantics to adapt to the complexity of the detection object sizes and shapes. URM introduces edge-guided feature learning to obtain richer textures and reduce the interference of irrelevant objects. Extensive experiments on three RSI-SOD datasets validate the superiority of our proposal.
python 3.8 + pytorch 1.9.0
We provide saliency maps (code: qyqq) on ORSSD, EORSSD, and ORSI4199 datasets. In addition, we also provide measure results (.mat) (code: qyqq) on the three datasets.
Download res2net50_v1b_26w_4s.pth (code: qyqq), and put it in './pre/'.
You can use the evaluation tool (MATLAB version) to evaluate the above saliency maps.