Research map / Point clouds / Segmentation and detection
Density-guided Translator Boosts Synthetic-to-Real Unsupervised Domain Adaptive Segmentation of 3D Point Clouds
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 non-learnable density-guided translator (DGT) that discards LiDAR points to match target point density, integrated into a two-stage self-training pipeline (DGT-ST) with prototype-based category-level adversarial warm-up (PCAN) and source-aware consistency LaserMix (SAC-LM) for synthetic-to-real 3D UDA segmentation.
Problem
Synthetic-to-real unsupervised domain adaptive semantic segmentation of LiDAR point clouds is hurt by sensor sampling-pattern (point density) mismatch and by weak initialization of self-training models.
Approach
A statistics-based, non-learnable translator that partitions scans by distance to the LiDAR center and randomly discards points per area to match the other domain's density, applied at the input level; combined with prototype-guided category-level adversarial warm-up and a source-aware consistency term in LaserMix-based self-training.
Key results
- DGT-ST reaches 43.1% mIoU on SynLiDAR→SemanticKITTI, +22.7 over source only and +9.4 over LaserMix-based self-training baseline (per abstract).
- DGT-ST reaches 50.8% mIoU on SynLiDAR→SemanticPOSS, +12.5 over source only and +4.3 over the best compared method PMAN (46.5).
- Density-guided discarding gives +1.2% over the baseline in PCAN on Syn→Sk, while random discarding drops mIoU by 0.9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGT in DGT-ST rather than LaserMix alone (ablation) | LaserMix self-training on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 36.0 -> 37.5 | Helped |
| Use density-guided point discarding instead of random point discarding (ablation) | PCAN on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 34.9 -> 37.0 | Helped |
| Use random point discarding (ablation) | PCAN with XY-noise on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 35.8 -> 34.9 | Hurt |
| Add XY noise (ablation) | PCAN on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 35.4 -> 35.8 | Helped |
| Use PCAN category-level adversarial alignment (ablation) | Adversarial adaptation on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 30.5 -> 35.4 | Helped |
| Add source-aware consistency (SAC) (ablation) | LaserMix self-training on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 37.5 -> 38.7 | Helped |
| Initialize self-training with the PCAN model (ablation) | LaserMix on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 36.0 -> 39.8 | Helped |
| Add beam discarding (ablation) | PCAN on SynLiDAR → SemanticPOSS · SynLiDAR → SemanticPOSS | mIoU: 42.3 -> 43.6 | Helped |
| Add point discarding after beam discarding (ablation) | PCAN on SynLiDAR → SemanticPOSS · SynLiDAR → SemanticPOSS | mIoU: 43.6 -> 44.4 | Helped |
| Add beam and point discarding (ablation) | LaserMix on SynLiDAR → SemanticPOSS · SynLiDAR → SemanticPOSS | mIoU: 45.5 -> 47.7 | Helped |
| Add DGT (ablation) | CoSMix on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 29.9 -> 30.5 | Helped |
| Initialize CoSMix with the PCAN model (ablation) | CoSMix + DGT on SynLiDAR → SemanticKITTI · SynLiDAR → SemanticKITTI | mIoU: 30.5 -> 39.1 | Helped |
Surprising findings
- Random point discarding reduces performance (drops mIoU by 0.9% in an ablation), indicating that naive density matching can enlarge the domain gap.
Limitations
- Discarding points only reduces density and cannot add points or fix noise; noise handling is limited to XY jitter.
- Results reported on only two synthetic-to-real tasks.
- Hyperparameters such as gamma2 and the number of partitions are tuned on the target validation set.
Experimental setup
Datasets: SynLiDAR SemanticKITTI SemanticPOSS
Benchmarks: SynLiDAR→SemanticKITTI SynLiDAR→SemanticPOSS
Metrics: mIoU
Baselines: AdaptSegNet CLAN ADVENT FADA MRNet PMAN CoSMix PolarMix LaserMix Source only
Builds on: LaserMix CoSMix ADVENT mean teacher LiDAR-Distillation
Closest papers by meaning
- Learning From Synthetic Data: Addressing Domain Shift for Semantic Segmentation (CVPR 2018 · similarity 0.89)
- Learning Semantic Segmentation From Synthetic Data: A Geometrically Guided Input-Output Adaptation Approach (CVPR 2019 · similarity 0.89)
- Adversarially Masking Synthetic To Mimic Real: Adaptive Noise Injection for Point Cloud Segmentation Adaptation (CVPR 2023 · similarity 0.89)
- Domain Adaptation on Point Clouds via Geometry-Aware Implicits (CVPR 2022 · similarity 0.88)
- Self-Supervised Augmentation Consistency for Adapting Semantic Segmentation (CVPR 2021 · similarity 0.88)
- Density Matters: Improved Core-Set for Active Domain Adaptive Segmentation (AAAI 2024 · similarity 0.86)
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.