Research map / Label-efficient and robust learning / Domain adaptation and generalization
Omni-Query Active Learning for Source-Free Domain Adaptive Cross-Modality 3D 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 OmniQuery, an active learning method for source-free cross-modal (2D image + 3D LiDAR) 3D semantic segmentation that decides when to query (curve-fitting Query Decider), which points to label (feature-ambiguity Point Ranker), and where to allocate budget (SAM-based Budget Slicer).
Problem
Active learning for source-free domain adaptive cross-modality 3D semantic segmentation: choosing when to query labels during self-training, which points are most informative, and how to spread the annotation budget spatially for diversity under a small label budget.
Approach
Combines three modules: fits logarithmic curves to pseudo-label mIoU during training to time queries before noise memorization (Early Learning), ranks individual points by entropy of neighboring-class proportions in feature space, and allocates per-region labeling quotas using SAM masks prompted with box prompts derived from the model's own predictions.
Key results
- With a 3% label budget, OmniQuery reaches 99.64% of fully supervised 2D+3D mIoU on USA→Singapore and 92.78% and 91.06% on the two larger-gap scenarios.
- Compared with the Annotator baseline, OmniQuery improves 2D+3D mIoU by 2.29%, 1.90%, 2.72%, and 2.95% across the four scenarios.
- Ablation on USA→Singapore shows Point Ranker, Query Decider, and Budget Slicer each add gains, totaling 2.45% 2D+3D mIoU over random selection.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OmniQuery instead of Annotator | Cross-modal SFDA segmentation · USA→Singapore, 3% labels | 2D+3D mIoU (%): 77.16 -> 79.45 | Helped |
| Use OmniQuery instead of Random | Cross-modal SFDA segmentation · Day→Night, 3% labels | 2D+3D mIoU (%): 68.64 -> 70.08 | Helped |
| Use OmniQuery instead of Margin | Cross-modal SFDA segmentation · VirtualKITTI→SemanticKITTI, 3% labels | 2D+3D mIoU (%): 68.73 -> 70.78 | Helped |
| Use OmniQuery instead of Random | Cross-modal SFDA segmentation · A2D2→SemanticKITTI, 3% labels | 2D+3D mIoU (%): 58.36 -> 60.01 | Helped |
| Use OmniQuery instead of Random | Cross-modal SFDA segmentation · USA→Singapore, 1% labels | 2D+3D mIoU (%): 75.85 -> 78.76 | Helped |
| Add Point Ranker to random point selection (ablation) | OmniQuery · USA→Singapore, 3% labels | 2D+3D mIoU (%): 77.00 -> 78.32 | Helped |
| Add Query Decider to Point Ranker (ablation) | OmniQuery · USA→Singapore, 3% labels | 2D+3D mIoU (%): 78.32 -> 79.02 | Helped |
| Add Budget Slicer to Point Ranker (ablation) | OmniQuery · USA→Singapore, 3% labels | 2D+3D mIoU (%): 78.32 -> 79.01 | Helped |
| Use all three modules instead of random selection (ablation) | OmniQuery · USA→Singapore, 3% labels | 2D+3D mIoU (%): 77.00 -> 79.45 | Helped |
| Use prompted SAM instead of SLIC for region segmentation (ablation) | OmniQuery Budget Slicer · USA→Singapore, 3% labels | 2D+3D mIoU (%): 76.91 -> 79.45 | Helped |
| Use prompted SAM instead of vanilla SAM (ablation) | OmniQuery Budget Slicer · USA→Singapore, 3% labels | 2D+3D mIoU (%): 79.11 -> 79.45 | Helped |
| Use ViT huge rather than ViT base or ViT large SAM (ablation) | OmniQuery Budget Slicer · USA→Singapore, 3% labels | 2D+3D mIoU (%): 79.29 -> 79.45 | Helped |
Surprising findings
- On the Day→Night scenario, most active methods outperform the fully supervised Oracle, which the authors attribute to extra supervision damaging source-model knowledge.
Limitations
- Oracle annotations are simulated by using ground-truth labels from the target training set rather than real annotators.
- Query timing relies on a fitted logarithmic curve and a manually chosen threshold tau (0.85) and K (4) hyperparameters.
- Only evaluated on LiDAR-camera driving datasets with four scenarios.
Experimental setup
Datasets: nuScenes-Lidarseg VirtualKITTI SemanticKITTI A2D2
Benchmarks: nuScenes-Lidarseg USA→Singapore nuScenes-Lidarseg Day→Night VirtualKITTI→SemanticKITTI A2D2→SemanticKITTI
Metrics: mIoU 2D mIoU 3D mIoU 2D+3D mIoU
Baselines: Random Confidence Entropy Margin ReDAL Annotator ST (self-training)
Builds on: Annotator ReDAL SAM Cross-modal UDA (xMUDA) SparseConvNet
Closest papers by meaning
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.86)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.86)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.85)
- Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation (CVPR 2024 · similarity 0.85)
- AdaCo: Overcoming Visual Foundation Model Noise in 3D Semantic Segmentation via Adaptive Label Correction (AAAI 2025 · similarity 0.85)
- OA-CNNs: Omni-Adaptive Sparse CNNs for 3D Semantic Segmentation (CVPR 2024 · 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.