Research map / Label-efficient and robust learning / Domain adaptation and generalization

Parsing All Adverse Scenes: Severity-Aware Semantic Segmentation with Mask-Enhanced Cross-Domain Consistency

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Domain-generalized segmentation · 13 citations (OpenAlex)

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 PASS, a unified unsupervised domain adaptation model for semantic segmentation across all adverse scenes (fog, night, rain, snow) using style augmentation with image merging, severity-aware masking, and SSIM-based cross-domain consistency analysis.

Problem

Unsupervised domain adaptation for semantic segmentation performs well on rain and snow but weakly on fog and night; prior methods over-focus on weather-specific features, and no single model handles all adverse scenes.

Approach

Augment target adverse images with style transfer (SA) merged with originals (MSA) to remove weather-specific features, then apply a severity-gated random photometric mask to source images (SPM) so that high-severity target batches are aligned with masked source images and low-severity ones with unmasked, with cross-domain consistency quantified by SSIM.

Key results

• On Cityscapes to ACDC-All, PASS reaches 70.8 mIoU versus 68.0 for HRDA; the ablation shows MSA alone gives +0.7 and SPM with MSA gives +2.8 overall.
• On Cityscapes to Foggy Zurich, PASS reaches 59.9 mIoU, a 9.3-point gain over FogAdapt (50.6).
• On Cityscapes to ACDC-Night, PASS reaches 60.2 mIoU, 7.3 points above HRDA (55.9 on night-scene comparison).

Measured effects reported in the paper

Surprising findings

• Plain style augmentation (SA) alone performs about 4 mIoU worse than the HRDA baseline, attributed to loss of content information that HRDA is sensitive to.
• A single random mask operation without severity perception drops performance by 5.8 mIoU versus HRDA.

Limitations

• Relies on the HRDA high-resolution pipeline, whose sensitivity to content information motivated the merging step.
• Night-scene evaluation on ACDC is reported on a subset (106 night images in validation) and some benchmarks use coarse labels (Nighttime Driving), which limits comparability.

Experimental setup

Datasets: Cityscapes ACDC Foggy Zurich Dark Zurich Painter by Numbers (style transfer training)

Benchmarks: ACDC-Fog Foggy Zurich Foggy Driving ACDC-Night Dark Zurich Nighttime Driving BDD100K-Night ACDC-Rain ACDC-Snow ACDC-All

Metrics: mIoU SSIM

Baselines: SFSU CMAda3 CuDA-Net FIFO FogAdapt GCMA MCGDA DANNet AdaptSeg DAFormer SePiCo HRDA

Builds on: HRDA DAFormer DACS ClassMix Mean Teacher

Closest papers by meaning

• Semantic Segmentation in Multiple Adverse Weather Conditions with Domain Knowledge Retention (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 0.90)
• 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 0.89)
• No Thing, Nothing: Highlighting Safety-Critical Classes for Robust LiDAR Semantic Segmentation in Adverse Weather (CVPR 2025 · similarity 0.88)
• Semantic Segmentation on Raindrop Degraded Images Using Two-Stage Dual Teacher-Student Learning (AAAI 2025 · similarity 0.86)