Reproduce “FSANet: Frequency-Spatial Aware Network for Image 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
Ruibo Wang, Ziyi Shen, Huaming Wu, Dong Liang, Kun Shang
FSANet adapts a frozen Segment Anything Model for segmentation in cluttered, degraded, and fine-structured images by combining structural priors, adaptive frequency filtering, spatial enhancement, and explicit edge supervision. Across five fine-grained benchmarks, its variants reach up to 89.0% mIoU and 81.4% mBIoU, while adding roughly 0.035 million trainable parameters and 2 GFLOPs.
Key insights
- The channel-wise RGB range provides a lightweight structural cue that helps preserve object shape and supports boundary prediction under changing illumination.
- Adaptive frequency processing complements spatial modeling: the learned filter can restore detail in blurred or hazy images while suppressing degradation-related spectral noise.
- Contour-focused evaluation exposes gains that region-overlap metrics alone may miss; the strongest improvements occur on thin and irregular objects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.