Research map / Label-efficient and robust learning / Domain adaptation and generalization
BEV-DG: Cross-Modal Learning under Bird's-Eye View for Domain Generalization of 3D Semantic Segmentation
Miaoyu Li, Yachao Zhang, Xu Ma, Yanyun Qu, Yun Fu
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 BEV-based area-to-area fusion (BAF) to baseline (ablation) | Baseline · A,N->S (SemanticKITTI) | mIoU Avg: 44.7 -> 52.9 | Improved |
| add BEV-driven domain contrastive learning (BDCL) to baseline (ablation) | Baseline · A,N->S (SemanticKITTI) | mIoU Avg: 44.7 -> 49.4 | Improved |
| combine BAF and BDCL (ablation) | Baseline · A,S->N (nuScenes) | mIoU Avg: 53.9 -> 59.0 | Improved |
| add BEV transformation (use BEV areas rather than point-level concatenation) (ablation) | BEV-DG · A,N->S (SemanticKITTI) | mIoU Avg: 38.7 -> 60.2 | Improved |
| add fusion of 2D and 3D BEV feature maps (ablation) | BEV-DG · A,N->S (SemanticKITTI) | mIoU Avg: 45.3 -> 60.2 | Improved |
| add Density-maintained Vector Modeling (DVM) (ablation) | BEV-DG · A,N->S (SemanticKITTI) | mIoU Avg: 50.0 -> 60.2 | Improved |
| add Density Transfer (DT) for positive pairs (ablation) | BEV-DG · A,N->S (SemanticKITTI) | mIoU Avg: 49.1 -> 60.2 | Improved |
| area-to-area fusion instead of point-to-point cross-modal learning under projection misalignment | BAF / BEV-DG vs xMUDA, Dual-Cross · A,S->N (nuScenes) | mIoU Avg | Improved |
| vary BEV area size w from 0.05m to 1m (ablation) | BEV-DG · A,S->N (nuScenes) | mIoU | Mixed |
| vary contrastive loss weight from 0.001 to 0.1 (ablation) | BEV-DG · A,S->N (nuScenes) | mIoU | No clear effect |
| BEV-DG versus best prior method on 3D mIoU in A,N->S | BEV-DG · A,N->S (SemanticKITTI) | mIoU 3D: 56.5 -> 54.7 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- xMUDA — xMUDA: Cross-Modal Unsupervised Domain Adaptation for 3D Semantic Segmentation (CVPR 2020) · better on 1 of 1 shared result
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
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.89)
- EAGLE: Efficient Adaptive Geometry-based Learning in Cross-view Understanding (NeurIPS 2024 · similarity 0.88)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.87)
- Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap (ICLR 2025 · similarity 0.87)
- Collaborating Foundation Models for Domain Generalized Semantic Segmentation (CVPR 2024 · similarity 0.87)
- Sparse All-Layer Connector for Domain Generalised Semantic Segmentation (NeurIPS 2026 · similarity 0.87)
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.
