Research map / Label-efficient and robust learning / Domain adaptation and generalization
Semantics Distortion and Style Matter: Towards Source-free UDA for Panoramic 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 a source-free UDA framework for pinhole-to-panoramic semantic segmentation that uses tangent projection and fixed-FoV ERP patches to extract source-model knowledge, then transfers it via panoramic prototype adaptation and cross-dual attention.
Problem
Source-free unsupervised domain adaptation for semantic segmentation of panoramic (ERP) images using only a pinhole-trained source model and unlabeled panoramic images, under semantic mismatch (FoV), distortion, and style discrepancies.
Approach
Extracts knowledge from the pinhole source model using tangent projection and fixed-FoV perspective patches of ERP images, integrates them into a globally updated panoramic prototype for adaptation, and aligns ERP and FFP features with a BN-statistics-guided cross-dual attention (CDAM) loss.
Key results
- On Cityscapes-to-DensePASS, the method reaches 50.12% mIoU with SegFormer-B2, +11.47% over the source model, outperforming SFDA and DTAC.
- On SynPASS-to-DensePASS (13 classes), mIoU reaches 42.18% with SegFormer-B2, +6.37% over the source baseline.
- On Stanford2D3D indoor, mIoU reaches 65.75% with SegFormer-B2, +10.99% over SFDA-style baseline and close to MPA with source data (67.16%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed source-free adaptation method instead of the strongest listed SFDA baseline | SegFormer-B1 · SynPASS-to-DensePASS | mIoU: 38.54 -> 41.78 | Helped |
| Use the proposed source-free adaptation method instead of DTAC | SegFormer-B1 · Cityscapes-to-DensePASS | mIoU: 43.06 -> 48.78 | Helped |
| Use the proposed method with b2 instead of b1 | Stanford2D3D panoramic · Stanford2D3D panoramic | mIoU: 57.63 -> 65.75 | Helped |
| Add prototype adaptation loss (Lppa) to prediction supervision (ablation) | Proposed method · Cityscapes-to-DensePASS | mIoU: 38.65 -> 45.42 | Helped |
| Add source-model fine-tuning loss (Lsft) to Lsup + Lppa (ablation) | Proposed method · Cityscapes-to-DensePASS | mIoU: 45.42 -> 46.23 | Helped |
| Add cross-dual attention loss (Lcda) to prediction supervision (ablation) | Proposed method · Cityscapes-to-DensePASS | mIoU: 38.65 -> 44.24 | Helped |
| Add BNS-guided feature loss (Lbns) to prediction supervision (ablation) | Proposed method · Cityscapes-to-DensePASS | mIoU: 38.65 -> 44.79 | Helped |
| Combine all proposed losses (ablation) | Proposed method · Cityscapes-to-DensePASS | mIoU: 38.65 -> 48.78 | Helped |
| Combine TP and FFP prototypes instead of using either projection alone (ablation) | Prototype adaptation · Cityscapes-to-DensePASS | mIoU: 44.14 -> 45.42; 44.28 -> 45.42 | Helped |
| Set FFP field of view to 90° rather than removing fixed FoV (ablation) | FFP · Cityscapes-to-DensePASS | mIoU: 38.65 -> 44.28 | Helped |
| Increase FFP field of view from 90° to 180° (ablation) | FFP · Cityscapes-to-DensePASS | mIoU: 44.28 -> 41.65 | Hurt |
| Increase FFP field of view from 90° to 360° (ablation) | FFP · Cityscapes-to-DensePASS | mIoU: 44.28 -> 40.31 | Hurt |
Surprising findings
- Naively applying existing pinhole SFUDA methods yields only limited gains on panoramic segmentation.
- The source-free method is comparable to some UDA methods that use source pinhole data.
Limitations
- Tangent projection during training adds computational cost.
- Segmentation performance still has room for improvement.
- Semantic mismatch between pinhole and panoramic domains is not fully addressed by current components.
Experimental setup
Datasets: Cityscapes DensePASS SynPASS Stanford2D3D
Benchmarks: DensePASS SynPASS Stanford2D3D
Metrics: mIoU
Baselines: SFDA DATC ProDA GTA HCL Simt MPA Trans4PASS DPPASS DATR DAFormer
Builds on: SFDA (dual attention source-free segmentation) MPA (mutual prototype adaptation) Trans4PASS
Closest papers by meaning
- GoodSAM: Bridging Domain and Capacity Gaps via Segment Anything Model for Distortion-aware Panoramic Semantic Segmentation (CVPR 2024 · similarity 0.92)
- Geometric Exploitation for Indoor Panoramic Semantic Segmentation (NeurIPS 2024 · similarity 0.87)
- AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation (NeurIPS 2026 · similarity 0.86)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.86)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.85)
- PanoWorld: Towards Spatial Supersensing in 360◦ Panorama World (NeurIPS 2026 · similarity 0.85)
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.