Research map / Detection and segmentation / Open-set and few-shot detection
Semantic Segmentation on Raindrop Degraded Images Using Two-Stage Dual Teacher-Student Learning
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 a two-stage dual teacher-student unsupervised domain adaptation method with raindrop-specific augmentation and mask-based recovery for semantic segmentation of raindrop-degraded images.
Problem
Semantic segmentation models degrade badly on images with adherent raindrops on lenses or windshields, which both corrupt appearance and cause misleading occlusion, and no target labels are available.
Approach
Decomposes raindrop degradation into visual-appearance degradation and raindrop occlusion, handled by two teacher-student pairs (a standard-augmentation reference pair and a physics-based raindrop-augmentation pair) with ensembled pseudo-labels, a mask-based recovery step that replaces low-confidence pseudo-labels using masked-image predictions, and a second stage that swaps augmentations between
Key results
- Achieves 60.3 mIoU on Rainy WCity, +4.4 mIoU over the best prior method (MIC at 55.9).
- Achieves 72.8 mIoU on ACDC Rainy, +2.3 mIoU over MIC (70.5).
- Ablation shows a single teacher-student pair with mixed augmentations performs worse than the MIC-style baseline, while the dual setup improves results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the two-stage dual teacher-student method | DAFormer · Rainy WCity | mIoU: +4.4 mIoU | Helped |
| use the two-stage dual teacher-student method | DAFormer · ACDC Rainy (Val.) | mIoU: +2.3 mIoU | Helped |
| use one teacher-student pair for standard and raindrop augmentations instead of separate pairs (ablation) | the proposed method · Rainy WCity | mIoU | Hurt |
| use the dual teacher-student framework with two raindrop types (ablation) | the proposed method · Rainy WCity | mIoU | Helped |
| replace fixed α with adaptive α (ablation) | Dual TS(2) · Rainy WCity | mIoU: approximate 1 mIoU improvement | Helped |
| add a third raindrop type (16 × 16) (ablation) | Dual TS(2a) · Rainy WCity | mIoU | No clear effect |
| add a fourth raindrop type (8 × 8) (ablation) | Dual TS(3a) · Rainy WCity | mIoU | No clear effect |
| apply mask-based recovery without updating the teacher-student pairs (ablation) | Dual TS(4a) and No Dual TS(1) · Rainy WCity | mIoU | Hurt |
| add two-stage updating without mask-based recovery (ablation) | Dual TS(4a) · Rainy WCity | mIoU | Helped |
| add mask-based recovery to two-stage updating (ablation) | the proposed method · Rainy WCity | mIoU | Helped |
Surprising findings
- Merging standard and raindrop augmentations into a single teacher-student pair degrades performance below the standard-augmentation baseline.
- Directly applying mask-based recovery without first updating the teacher-student pairs causes a slight performance drop.
Limitations
- Relies on a physics-based raindrop model with fixed hand-chosen sizes and ratios for augmentation.
- Performance gains among later ablation configurations (Dual TS(2a)-(4a)) are minimal.
- Mask-based recovery benefit depends on the quality of the first-stage reference pair.
Experimental setup
Datasets: Cityscapes Rainy WCity ACDC Rainy Raindrop dataset
Benchmarks: Rainy WCity ACDC Rainy
Metrics: mIoU
Baselines: Restormer + DAFormer Weather Diffusion + DAFormer DAFormer HRDA MIC
Builds on: MIC DAFormer HRDA EMA teacher-student self-training
Closest papers by meaning
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.88)
- Towards Universal Rainy Image Restoration: Benchmark and Baseline (AAAI 2025 · similarity 0.88)
- ERF: A Benchmark Dataset for Robust Semantic Segmentation Under Extreme Rainfall Conditions (AAAI 2025 · similarity 0.87)
- Boosting Image De-Raining via Central-Surrounding Synergistic Convolution (AAAI 2025 · similarity 0.86)
- Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (AAAI 2024 · similarity 0.86)
- End-to-End Video Semantic Segmentation in Adverse Weather using Fusion Blocks and Temporal-Spatial Teacher-Student Learning (NeurIPS 2024 · similarity 0.86)
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.