Research map / Label-efficient and robust learning / Noisy labels
CoSW: Conditional Sample Weighting for Smoke Segmentation with Label Noise
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 conditional sample weighting (CoSW) with a regularized within-prototype entropy to handle label noise in smoke segmentation, evaluated on real-noisy SmokeSeg/SMOKE5K and a synthetic noisy NS-1K dataset.
Problem
Pixel-level smoke segmentation datasets contain noisy labels due to blurry, non-rigid smoke edges and variable transparency, which existing noisy-label methods (designed for classification with i.i.d. noise) do not handle.
Approach
Uses a multi-prototype framework where prototypes act as priors so each feature cluster gets its own weighting criterion; a regularized within-prototype entropy (maximum-entropy objective with an L2 penalty) yields closed-form per-pixel weights and a weighted prototype update, plus a regularized scatter metric loss.
Key results
- On SmokeSeg with MiT-B3, CoSW reaches 72.32 F1 / 59.25 mIoU versus 70.17 / 56.94 for CleanNet and 67.70 / 53.37 for SegFormer.
- On SMOKE5K, the method achieves 81.71 F-beta and 71.24 mIoU, above CleanNet's 80.37 / 70.23.
- On synthetic NS-1K at 80% high-intensity noise, F1 is 42.37 vs 37.38 for Trans-BVM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoSW sample weighting to prototypes (ablation) | SegFormer on SmokeSeg · SmokeSeg | F1: 68.17 -> 70.04 | Helped |
| add CoSW sample weighting to prototypes (ablation) | SegFormer on SmokeSeg · SmokeSeg | mIoU: 55.23 -> 56.40 | Helped |
| add CoSW weighting to prototype updates (ablation) | prototype-based SegFormer on SmokeSeg · SmokeSeg | F1: 68.17 -> 69.38 | Helped |
| add CoSW weighting to prototype updates (ablation) | prototype-based SegFormer on SmokeSeg · SmokeSeg | mIoU: 55.23 -> 55.72 | Helped |
| combine sample weighting and prototype-update weighting (ablation) | prototype-based SegFormer on SmokeSeg · SmokeSeg | F1: 68.17 -> 71.39 | Helped |
| combine sample weighting and prototype-update weighting (ablation) | prototype-based SegFormer on SmokeSeg · SmokeSeg | mIoU: 53.37 -> 57.68 | Helped |
| use Triplet Loss instead of prototype-only metric learning (ablation) | prototype-based method on SmokeSeg · SmokeSeg | F1: 68.17 -> 67.80 | Hurt |
| add CoSW to Triplet Loss metric learning (ablation) | prototype-based method on SmokeSeg · SmokeSeg | F1: 67.80 -> 71.88 | Helped |
| replace Shannon entropy with Kapur’s entropy (ablation) | CoSW on SmokeSeg · SmokeSeg | F1: 72.32 -> 71.34 | Hurt |
| replace Shannon entropy with Burg’s entropy (ablation) | CoSW on SmokeSeg · SmokeSeg | F1: 72.32 -> 69.20 | Hurt |
| increase Sinkhorn iterations from 1 to 3 (ablation) | CoSW pixel-prototype matching on SmokeSeg · SmokeSeg | F1: 66.40 -> 72.32 | Helped |
| increase Sinkhorn iterations from 3 to 20 (ablation) | CoSW pixel-prototype matching on SmokeSeg · SmokeSeg | F1: 72.32 -> 72.10 | Hurt |
Surprising findings
- Sample weighting with CoSW on its own gives a large gain, and prototype-based decoding alone already improves over SegFormer.
- Triplet-loss metric learning under noisy labels reduces performance relative to the plain prototype model.
Limitations
- Relies on the assumption that most pixels have clean labels; with very high noise the prototypes can learn noise features.
- No clean validation set is used to verify prototype cleanliness.
- No error bars or statistical significance reported; compute resources not stated.
Experimental setup
Datasets: SmokeSeg SMOKE5K NS-1K
Benchmarks: SmokeSeg SMOKE5K NS-1K
Metrics: F1 mIoU
Baselines: AFFormer SeaFormer SegFormer SC CleanNet DeeplabV3+ OCRNet SegNeXt Trans-BVM
Builds on: CleanNet ProtoSeg SwAV M-estimation DeepLDA
Closest papers by meaning
- FoSp: Focus and Separation Network for Early Smoke Segmentation (AAAI 2024 · similarity 0.87)
- Stratify or Die: Rethinking Data Splits in Image Segmentation (NeurIPS 2025 · similarity 0.81)
- Label-Noise Robust Diffusion Models (ICLR 2024 · similarity 0.81)
- L2B: Learning to Bootstrap Robust Models for Combating Label Noise (CVPR 2024 · similarity 0.81)
- Enhanced Sample Selection with Confidence Tracking: Identifying Correctly Labeled Yet Hard-to-Learn Samples in Noisy Data (AAAI 2025 · similarity 0.81)
- Learning from Noisy Labels via Conditional Distributionally Robust Optimization (NeurIPS 2024 · similarity 0.81)
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.