Research map / Label-efficient and robust learning / Domain adaptation and generalization
Balanced Learning for Domain Adaptive Semantic Segmentation
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 BLDA, a plug-and-play logits-distribution alignment method (post-hoc and online GMM-based) with a cumulative-density regression auxiliary loss to reduce class bias in self-training UDA for semantic segmentation.
Problem
Self-training for unsupervised domain adaptive semantic segmentation learns classes unevenly due to class imbalance and label/data distribution shift between domains, causing over- and under-predicted classes.
Approach
Assess class bias from per-class positive and negative logit distributions, align each class's logit CDF to shared anchor distributions (post-hoc and online via GMMs with logit correction terms in the loss), and use the resulting cumulative density value as a domain-shared auxiliary regression target.
Key results
- On GTA->CS with DAFormer (MiT-B5), BLDA raises mIoU from 68.3 to 70.7 (ablation, with CDE), and integrated into MIC it reaches 75.9 to 77.1 mIoU per Table 1.
- mAcc improvements range from 2.9% to 5.6% across baselines on GTA->CS.
- BLDA with HRDA on SYNTHIA->CS increases mIoU from 65.8 to 67.9 (16-class mIoU).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BLDA to DACS | DACS · GTA→CS | mIoU: 52.1 -> 54.7 | Helped |
| Add BLDA to DAFormer | DAFormer · GTA→CS | mIoU: 68.3 -> 70.7 | Helped |
| Add BLDA to MIC | MIC · SYN→CS | mIoU: 67.3 -> 69.1 | Helped |
| Add BLDA to DAFormer | DAFormer · GTA→CS | mAcc: 77.8 -> 82.0 | Helped |
| Add post-hoc logit adjustment (ablation) | DAFormer on GTA→CS · GTA→CS | mIoU: 68.3 -> 69.2 | Helped |
| Add source-only online logit adjustment (ablation) | DAFormer on GTA→CS · GTA→CS | mIoU: 68.3 -> 68.9 | Helped |
| Apply online logit adjustment in source and target domains (ablation) | DAFormer on GTA→CS · GTA→CS | mIoU: 68.3 -> 70.2 | Helped |
| Add cumulative density estimation to OLA (ablation) | DAFormer on GTA→CS · GTA→CS | mIoU: 70.2 -> 70.7 | Helped |
| Set GMM components K to 5 instead of K=1 (ablation) | BLDA with DAFormer · GTA→CS | mIoU: 70.2 -> 70.7 | Helped |
| Set GMM components K to 10 instead of K=5 (ablation) | BLDA with DAFormer · GTA→CS | mIoU: 70.7 -> 70.6 | Hurt |
| Set momentum factor to 0.99 (ablation) | BLDA with DAFormer · GTA→CS | mIoU: 68.6 -> 70.7 | Helped |
| Set EM loops to 3 rather than 1 (ablation) | BLDA with DAFormer · GTA→CS | mIoU: 70.4 -> 70.7 | Helped |
Surprising findings
- Maximizing mIoU does not align with maximizing mAcc: a balanced classifier (tau=1) boosts mAcc but a smaller tau is needed for best mIoU.
- Re-weighting alone gives unstable self-training gains, while re-sampling is more effective, according to the motivating experiments.
Limitations
- mIoU gains require reducing tau, which trades off against balanced accuracy (mAcc)
- Anchor distributions are estimated from source-domain global logits, an assumption that may not hold for other shifts
- Gains on some classes (e.g., train, bus) can decline slightly; results reported only for specific UDA frameworks
Experimental setup
Datasets: GTAv SYNTHIA Cityscapes ImageNet-1k VisDA-2017 VIPER BDD100k
Benchmarks: GTA->Cityscapes SYNTHIA->Cityscapes VisDA-2017 VIPER->Cityscapes-Seq VIPER->BDD100k
Metrics: mIoU mAcc std of per-class IoU/Acc
Baselines: DACS DAFormer CDAC HRDA MIC AdaptSegNet CRST PLCA ProDA CPSL TransDA ADFormer
Builds on: DAFormer HRDA MIC CDAC DACS
Closest papers by meaning
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.92)
- Class-Balanced Pixel-Level Self-Labeling for Domain Adaptive Semantic Segmentation (CVPR 2022 · similarity 0.92)
- Category Anchor-Guided Unsupervised Domain Adaptation for Semantic Segmentation (NeurIPS 2019 · similarity 0.92)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.91)
- Towards Unsupervised Domain Bridging via Image Degradation in Semantic Segmentation (NeurIPS 2025 · similarity 0.90)
- SemiDAViL: Semi-supervised Domain Adaptation with Vision-Language Guidance for Semantic Segmentation (CVPR 2025 · similarity 0.90)
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.