Reproduce “Frequency and Edge-Guided Segment Anything Model for Remote Sensing Image Semantic Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Feng Gao, Zizhe Pan, Haoting Wang, Ruzhuang Hua, Jingchao Cao, Junyu Dong, Qian Du
A framework called FE-SAM improves remote sensing image semantic segmentation by adapting the Segment Anything Model (SAM) to remote sensing images through frequency-domain modulation and edge-guided refinement. It achieves superior mIoU scores, such as 85.24% on ISPRS Vaihingen and 86.73% on ISPRS Potsdam, and enhances boundary precision by 2.72% on B-IoU (d=8) over the next best method.
Key insights
- Domain-aware architectural modifications, specifically frequency-domain adaptation, are effective in bridging the spectral gap between natural images (SAM's pre-training) and remote sensing data.
- Explicitly integrating multi-scale, edge-enhanced features extracted directly from the raw input image is crucial for recovering fine spatial details and achieving precise object boundary delineation.
- The synergistic combination of frequency-modulated feature adaptation and edge-guided refinement significantly enhances both the semantic accuracy and geometric precision of segmentation in remote sensing images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.