FSANet: Frequency-Spatial Aware Network for Image Segmentation
Ruibo Wang, Ziyi Shen, Huaming Wu, Dong Liang, Kun Shang
cs.CV
Quick verdict
SAM-based frequency-aware segmentation net plus SceneX dataset; gains broad and cheap, but annotation circularity and numeric inconsistencies need checking.
Summary
FSANet is a segmentation framework built on a frozen SAM (ViT-B/L/H) that adds three lightweight modules: a Structure Prior Module using a max-min channel prior with multi-kernel convolutions; a Dual-Domain Awareness Module applying learnable magnitude/phase FFT filters (CBAM-style) plus an SCConv spatial branch to fused encoder features; and an Edge Estimation Module supervised by an auxiliary boundary BCE loss (disabled at inference). The motivation is quantitative: per-degradation high-band energy (Table I) and image-to-feature spectrum correlations (+0.56 blur, -0.81 noise) justify filtering in feature space. All SAM variants are retrained under an identical data/schedule/prompt protocol and FSANet consistently beats HQ-SAM on SceneX, DIS5K-VD, COIFT, HRSOD, and ThinObject5K-TE (FSANet-H average mIoU 89.0 vs 87.9; mBIoU gaps of +1.3/+1.5/+2.4 widening with scale), at near-zero cost (+0.03M params, +2 GFLOPs, 1-3% slower). It reports the best zero-shot COCO AP (46.2) and 49.7 Mean AP on 25-task SegInW. The authors also release SceneX (11,000 images, 10 non-ideal scenarios) with semi-automated annotations. Main limitations: SceneX masks derive from SAM-family candidate models that are also baselines (circularity risk), unexplained numeric inconsistencies across tables, and thin statistical support for small margins.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Adding SPM, DDAM, and EEM (about 0.035M parameters, ~2 GFLOPs) to a frozen SAM consistently outperforms HQ-SAM across ViT-B/L/H on five fine-grained benchmarks, with boundary-mBIoU gains of +1.3/+1.5/+2.4 that widen with backbone capacity.
- Real-world noise has degradation-specific, disentangleable frequency signatures, and SAM's encoder feature spectrum inherits image-band statistics, which makes learnable spectral filtering of features (rather than images) meaningful.
- FSANet generalizes zero-shot, achieving the best COCO instance AP (46.2) among compared SAM variants and 49.7 Mean AP on the 25-task SegInW benchmark, above SAM (48.7) and HQ-SAM (49.6).
- SceneX (11,000 images, 10 non-ideal scenarios) is a robustness benchmark whose masks match manually annotated datasets on model-free shape statistics (boundary-gradient ratio 2.95 vs 3.55-6.36; thin-structure 0.182 within 0.168-0.369).
Strengths
Quantitative, falsifiable motivation for feature-level spectral filtering
Section II measures high-band energy under six controlled degradations (Table I: blur x0.04 to noise x2.71 relative to clean) and shows the frozen encoder's feature spectrum tracks image-band statistics selectively (blur +0.56, noise -0.81 correlation over 250 images), rather than asserting frequency priors axiomatically.
Broad evaluation with explicit cost accounting
Five fine-grained benchmarks plus zero-shot COCO and 25-task SegInW; Table III reports the modules cost +0.03M params, ~2 GFLOPs, unchanged peak memory, 1-3% FPS drop, and the FFT itself is measured at 0.34 GFLOPs and <0.5 ms.
Controlled comparison protocol
Dataset integrity audit for SceneX
Behavior-level interpretability of the learned filter
Weaknesses
Circularity risk in the SceneX benchmark
Unexplained numeric inconsistencies across tables
Statistical support is thin for the headline margins
Misleading efficiency framing on COCO
Overclaimed novelty
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FSANet and how does it differ from SAM?
FSANet is an image segmentation model built on a frozen Segment Anything Model (SAM). It adds three small modules: a structure-prior module computed from the max-min difference of RGB channels, a dual-domain module that filters the feature spectrum with learnable magnitude/phase FFT filters, and an edge-estimation branch trained with a boundary loss (turned off at inference). The additions cost roughly 0.035M parameters and ~2 GFLOPs, and the paper reports higher accuracy than HQ-SAM across ViT-B/L/H backbones.
What is the SceneX dataset?
SceneX is an open-source segmentation dataset introduced alongside FSANet: 11,000 images covering ten challenging non-ideal scenarios such as poor lighting and adverse weather, with resolutions from 200x125 to 4000x3000. Masks were produced semi-automatically - saliency seeding, five SAM-family candidate segmenters, best-IoU selection, OpenCV edge refinement, and manual validation - and both the automatic and manually refined subsets are released.
How much better is FSANet than HQ-SAM?
On five fine-grained benchmarks the paper reports average mIoU 89.0 vs 87.9 for the Huge versions (FSANet-H vs HQ-SAM-H), with boundary-mBIoU gaps of +1.3/+1.5/+2.4 for Base/Large/Huge. On zero-shot COCO it reports AP 46.2 vs 45.9 for HQ-SAM, and on the 25-task SegInW benchmark 49.7 vs 49.6 Mean AP - the latter two margins are small and reported without error bars.
Is FSANet's code and data available?
Yes. The paper states that code is available at github.com/wrbcode/FSANet and that the SceneX dataset is open-source, including both the automatic and manually refined annotation subsets; it also describes training details (12 epochs, lr 0.001, RTX A6000, PyTorch 1.13.1).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2, 3, 5, 6, and 8 carry key qualitative evidence (motivation visuals, architecture, dataset composition, qualitative comparisons, attention maps) but could not be assessed from the extracted text.
- Table VI's checkmark layout is garbled in extraction; ablation row attributions were reconstructed from the surrounding prose and may misattribute individual rows.
- Minor composition/extraction noise exists (e.g., unbalanced parenthesis in Eq. 13, the 'ten publicly available datasets' sentence whose 7.4k-masks description does not match the five listed evaluation sets); formatting judgments were avoided but some numbers may be extraction-damaged.
- The SceneX-annotation-circularity concern relies on the paper's own description of the annotation pipeline; the released manual-validation protocol may mitigate it in ways not visible from text alone.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.