Research map / Detection and segmentation / CNN features
FoSp: Focus and Separation Network for Early Smoke Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FoSp, a Focus and Separation network for early smoke segmentation with a bidirectional-cascade focus map, feature-level foreground/background separation via inpainters, and domain fusion, plus a new real-world SmokeSeg dataset.
Problem
Early smoke segmentation is hard because early smoke is small and transparent, causing high miss detection (low recall) and low precision in existing segmentation methods.
Approach
A bidirectional cascade generator guides low- and high-resolution features toward mid-resolution to produce a Focus Map covering the smoke scope; a Separation module uses two weight-shared inpainters and an L1 feature difference to extract a smoke foreground from a smoke-free background; a Domain Fusion MLP merges origin and foreground feature domains to balance recall and precision.
Key results
- On SmokeSeg, FoSp reaches F-beta 72.05% vs. SegFormer 67.99%, with small-smoke F-beta 66.03% (7.71 points above SegFormer).
- On SYN70K, FoSp achieves mIoU 83.00% on DS01, and on SMOKE5K it reaches F-beta 81.6% vs. Trans-BVM 79.1%.
- Ablation on SmokeSeg: adding the Focus module with Focus loss raises recall (71.77 to 75.60) and the Separation module raises precision (71.34 to 75.50), with the full model giving the best F-beta.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FoSp instead of Trans-BVM | SYN70K · SYN70K DS02 | mMse: 0.1362 -> 0.1081 | Helped |
| Use FoSp instead of Trans-BVM | SMOKE5K · SMOKE5K test | Fβ: 79.1 -> 81.6 | Helped |
| Use FoSp instead of SegFormer | SmokeSeg · SmokeSeg Small | Fβ: 58.32 -> 66.03 | Helped |
| Add Focus loss to the SegFormer baseline (ablation) | SmokeSeg · SmokeSeg Small | Recall: 60.82 -> 65.06 | Helped |
| Add Focus loss and the Focus module (ablation) | SegFormer baseline · SmokeSeg Small | Recall: 60.82 -> 68.52 | Helped |
| Add the Separation module (ablation) | SegFormer baseline with Focus loss and Focus · SmokeSeg Small | Precision: 62.86 -> 71.16 | Helped |
| Add Domain Fusion (ablation) | SegFormer baseline with Focus loss, Focus, and Separation · SmokeSeg | Fβ: 71.50 -> 72.05 | Helped |
| Use only foreground features for prediction (ablation) | FoSp · SmokeSeg, full test | Fβ: 67.99 -> 68.22 | Helped |
| Use image-level instead of feature-level separation (ablation) | FoSp · SmokeSeg, full test | Fβ: 72.05 -> 71.04 | Hurt |
| Set the Focus Map threshold to 0.9 instead of 0.15 (ablation) | FoSp · SmokeSeg, full test | Fβ: 72.05 -> 69.79 | Hurt |
Surprising findings
- A 'Foreground only' decoder variant already slightly outperforms SegFormer on F-beta (68.22 vs 67.99).
- Using the Focus Map alone as prediction at threshold 0.15 gives 88.74% recall, while FoSp still beats SegFormer at a high FM threshold of 0.9 (69.79 F-beta).
Limitations
- Image-level separation is computationally expensive, so the method uses feature-level separation, which is less transparency-precise.
- Separation relies on pre-trained inpainting network parameters and the Focus Map quality.
- Authors note no complete open-source smoke segmentation training code was available, so baselines were reimplemented.
Experimental setup
Datasets: SYN70K SMOKE5K SmokeSeg
Benchmarks: SYN70K SMOKE5K SmokeSeg
Metrics: mIoU F-beta MAE (M) recall precision
Baselines: SegFormer DeepLabv3+ OCRNet PSPNet FCN Trans-BVM EncNet CCNet
Builds on: SegFormer MiT-B3 Trans-BVM
Closest papers by meaning
- CoSW: Conditional Sample Weighting for Smoke Segmentation with Label Noise (NeurIPS 2024 · similarity 0.87)
- DetectiumFire: A Comprehensive Multi-modal Dataset Bridging Vision and Language for Fire Understanding (NeurIPS 2025 · similarity 0.81)
- Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · similarity 0.80)
- Guided Real Image Dehazing Using YCbCr Color Space (AAAI 2025 · similarity 0.80)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.79)
- EOV-Seg: Efficient Open-Vocabulary Panoptic Segmentation (AAAI 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.