Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unsupervised Domain Adaptation for Semantic Segmentation using Depth Distribution
Quanliang Wu, Huajun Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add density estimation branch to segmentation+depth model (ablation) | Multi-task UDA model · SYNTHIA->Cityscapes (16 classes) | mIoU: 41.7 -> 44.8 | Improved |
| Add confidence-based self-training (ablation) | Model M2 · SYNTHIA->Cityscapes (16 classes) | mIoU: 44.8 -> 47.6 | Improved |
| Add spatial prior pseudo-label refinement to self-training (ablation) | Model M3 · SYNTHIA->Cityscapes (16 classes) | mIoU: 47.6 -> 48.2 | Improved |
| Add branch balance loss with online density map on target domain (ablation) | Proposed model without self-training · SYNTHIA->Cityscapes (16 classes) | mIoU: 44.1 -> 44.8 | Improved |
| Use ground-truth depth rather than predicted depth to build source density map (ablation) | Proposed model without self-training · SYNTHIA->Cityscapes (16 classes) | mIoU: 43.4 -> 44.8 | Improved |
| Fuse all three branches' results instead of density+segmentation only (ablation) | Proposed model without self-training · SYNTHIA->Cityscapes (16 classes) | mIoU: 44.8 -> 37.8 | Hurt |
| Use DADA-style depth+segmentation fusion instead of density+segmentation fusion (ablation) | Proposed model without self-training · SYNTHIA->Cityscapes (16 classes) | mIoU: 44.8 -> 43.7 | Hurt |
| Add density estimation branch to improve depth (ablation) | M1 vs M2 · Cityscapes depth | |Rel|: 0.7 -> 0.5 | Improved |
| Add density estimation branch to improve depth accuracy (ablation) | M1 vs M2 · Cityscapes depth | delta1: 0.21 -> 0.26 | Improved |
| Raise auxiliary task weights to 1e-1 or lower them to 1e-3 (ablation) | Proposed model · SYNTHIA->Cityscapes (16 classes) | mIoU | Hurt |
| Vary auxiliary task weights within the 1e-2 level (ablation) | Proposed model · SYNTHIA->Cityscapes (16 classes) | mIoU | No clear effect |
| Use our model instead of CTRL | Proposed model · SYNTHIA->Mapillary (7 classes) 320x640 | mIoU: 62.5 -> 60.3 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DADA — DADA: Depth-Aware Domain Adaptation in Semantic Segmentation (ICCV 2019) · better on 6 of 6 shared results
- SPIGAN — SPIGAN: Privileged Adversarial Learning from Simulation (ICLR 2019) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Domain Adaptive Semantic Segmentation With Self-Supervised Depth Estimation (ICCV 2021 · similarity 0.91)
- Unsupervised Domain Adaptation for Semantic Segmentation via Class-Balanced Self-Training (ECCV 2018 · similarity 0.91)
- Unsupervised Model Adaptation for Continual Semantic Segmentation (AAAI 2021 · similarity 0.90)
- Balanced Learning for Domain Adaptive Semantic Segmentation (ICML 2025 · similarity 0.90)
- Geometric Unsupervised Domain Adaptation for Semantic Segmentation (ICCV 2021 · similarity 0.90)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.89)
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.
