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

Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation

NeurIPS 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Adaptation for semantic segmentation · 0 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 DiDA, a plug-and-play UDA plugin that builds continuous intermediate domains via forward-diffusion-style image degradation and compensates semantic shift with a diffusion encoder trained to reconstruct degraded inputs.

Problem

Self-training UDA for semantic segmentation overlooks explicit modeling of domain-shared feature extraction; naive degradation-based augmentation causes semantic shift (loss of discriminative features).

Approach

Treat incremental image degradation (Gaussian noise, blur, masking) as a diffusion forward process to create intermediate domains, and add a time-conditioned diffusion encoder with a reconstruction head whose residual features compensate the semantic shift, so the segmentation encoder learns domain-invariant features across degradation levels.

Key results

• On GTA->CS, DiDA on MIC reaches 76.8 mIoU (vs 75.5 reproduced MIC baseline), a new SOTA reported by the authors.
• DiDA on DAFormer improves GTA->CS from 68.3 to 70.3 mIoU (+2.0) and on CS->ACDC from 55.4 to 59.1 mIoU.
• Gains reported across DAFormer, HRDA, MIC (Transformer) and CNN-based MIC, ranging 0.8% to 3.7% mIoU.

Measured effects reported in the paper

Surprising findings

• Naive time-embedding-only degradation augmentation (g_time without compensation) performs heavily worse, attributed to semantic shift from excessive degradation.

Limitations

• Moderate training overhead from the diffusion encoder and extra forward passes on degraded inputs.
• Degradation operations and their combinations are manually chosen heuristically.
• Mask-based extension gives smaller gains than the default noise version.

Experimental setup

Datasets: GTAV SYNTHIA Cityscapes ACDC ImageNet-1k

Benchmarks: GTAV->Cityscapes SYNTHIA->Cityscapes Cityscapes->ACDC

Metrics: mIoU per-class IoU

Baselines: DAFormer HRDA MIC FST DiGA DTS ADFormer CoPT TransDA DACS ProDA DAP

Builds on: DACS DAFormer HRDA MIC Denoising Diffusion Probabilistic Models cold diffusion

Closest papers by meaning

• Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.91)
• Deliberated Domain Bridging for Domain Adaptive Semantic Segmentation (NeurIPS 2022 · similarity 0.91)
• Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer (AAAI 2021 · similarity 0.90)
• Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.90)
• Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.90)
• Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.90)