Research map / Continual learning / Domain and prompt-based
Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention
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 continual multi-target UDA method for semantic segmentation under sequential adverse weather (night, rain, fog, snow) using adaptive knowledge acquisition, pseudo-label blending with the previous teacher, and Fourier-based weather composition replay.
Problem
Semantic segmentation degrades under unlabeled adverse weather; sequentially adapting to multiple unlabeled weather domains causes forgetting of earlier knowledge while acquiring new knowledge.
Approach
Combines a previous-teacher-derived confidence and class-wise feature-distance weighting mask to suppress learning from harmful target regions, blends pseudo-labels from the previous and current teachers, and replays past weather information by injecting averaged Fourier amplitude spectra into random segments of new target images, without storing past target images.
Key results
- Ours (DAFormer) achieves 65.7 average mIoU and 3.6 accumulated forgetting on the Cityscapes→Night→Rain→Fog→Snow sequence, versus MIC's 60.1 and 11.3.
- Ours (DeepLabV2) reaches 40.2 average mIoU and 4.3 accumulated forgetting, beating MuHDi (33.9 mIoU, 18.4 A.F.).
- Without source data, the method reaches 63.0 average mIoU and 10.3 A.F., versus MIC's 57.5 and 23.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed method instead of MIC | DAFormer · ACDC four-target sequence | mIoU Avg. (%): 60.1 -> 65.7 | Helped |
| Use the proposed method instead of MIC | DAFormer · ACDC four-target sequence | A.F. (%): 11.3 -> 3.6 | Helped |
| Use the proposed method instead of MuHDi | DeeplabV2 · ACDC four-target sequence | mIoU Avg. (%): 33.9 -> 40.2 | Helped |
| Use the proposed method instead of MuHDi | DeeplabV2 · ACDC four-target sequence | A.F. (%): 18.4 -> 4.3 | Helped |
| Adapt without source data | Ours · ACDC four-target sequence | mIoU Avg. (%): 57.5 -> 63.0 | Helped |
| Adapt without source data | Ours · ACDC four-target sequence | A.F. (%): 23.9 -> 10.3 | Helped |
| Add pseudo-label blending (ablation) | AdvEnt · ACDC four-target sequence | mIoU Avg. (%): 32.4 -> 34.0 | Helped |
| Add pseudo-label blending (ablation) | AdvEnt · ACDC four-target sequence | A.F. (%): 22.4 -> 19.6 | Helped |
| Add model-level adjustment (ablation) | AdvEnt · ACDC four-target sequence | A.F. (%): 22.4 -> 12.9 | Helped |
| Add feature-level adjustment to model-level adjustment (ablation) | AdvEnt · ACDC four-target sequence | A.F. (%): 12.9 -> 9.6 | Helped |
| Add weather composition replay to model- and feature-level adjustment (ablation) | AdvEnt · ACDC four-target sequence | A.F. (%): 9.6 -> 4.5 | Helped |
Limitations
- Hyperparameters (alpha schedule, sigma range, patch size) are set empirically.
- Feature-level adjustment requires source data, so it is not applicable in the without-source setting.
- Continual setting is evaluated only on the ACDC four-target sequence.
Experimental setup
Datasets: Cityscapes ACDC
Benchmarks: ACDC
Metrics: mIoU accumulated forgetting
Baselines: MIC AdvEnt MuHDi MTKT Multi-Dis
Builds on: MIC AdvEnt DAFormer MuHDi
Closest papers by meaning
- Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency (AAAI 2024 · similarity 0.91)
- End-to-End Video Semantic Segmentation in Adverse Weather using Fusion Blocks and Temporal-Spatial Teacher-Student Learning (NeurIPS 2024 · similarity 0.91)
- UniMix: Towards Domain Adaptive and Generalizable LiDAR Semantic Segmentation in Adverse Weather (CVPR 2024 · similarity 0.89)
- ERF: A Benchmark Dataset for Robust Semantic Segmentation Under Extreme Rainfall Conditions (AAAI 2025 · similarity 0.88)
- Language-driven All-in-one Adverse Weather Removal (CVPR 2024 · similarity 0.87)
- Continuous Adverse Weather Removal via Degradation-Aware Distillation (CVPR 2025 · similarity 0.87)
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.