Research map / Label-efficient and robust learning / Semi-supervised learning
Exploring Scene Affinity for Semi-Supervised LiDAR 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 AIScene, a semi-supervised LiDAR segmentation framework that erases points without pseudo-labels to keep forward and backward propagation consistent and uses multi-scene patch/instance mixing augmentation.
Problem
Semi-supervised LiDAR semantic segmentation in driving scenes suffers from intra-scene inconsistency (unlabeled points used in forward but not backward propagation) and limited semantic diversity from two-scene mixing augmentations.
Approach
Remove points lacking pseudo-labels from pseudo-labeled scenes (point erasure) so forward and backward propagation are consistent, and mix patches and instances drawn from pools spanning multiple scenes (MixPatch and InsFill) to increase semantic diversity.
Key results
- On SemanticKITTI with 1% labels, AIScene reaches 61.2 mIoU vs 59.3 for DDSemi (+1.9); on nuScenes at 1% it reaches 60.2 vs 58.1 (+2.1).
- Average over four label ratios improves over prior SOTA by 1.3% on SemanticKITTI and 1.2% on nuScenes.
- Applied to Cylinder3D, AIScene raises SemanticKITTI average mIoU from 54.5 to 61.5 (+7.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AIScene instead of DDSemi | Semi-supervised LiDAR segmentation · SemanticKITTI, average across 1%, 10%, 20%, and 50% labeled data | mIoU: 64.4 -> 65.7 (+1.3%) | Helped |
| Use AIScene instead of DDSemi | Semi-supervised LiDAR segmentation · nuScenes, average across 1%, 10%, 20%, and 50% labeled data | mIoU: 69.7 -> 70.9 (+1.2%) | Helped |
| Use AIScene instead of DDSemi | Semi-supervised LiDAR segmentation · nuScenes 50% | mIoU: 76.5 -> 76.4 (-0.1) | Hurt |
| Add point erasure (ablation) | Cylinder3D semi-supervised baseline · SemanticKITTI, 1%, 10%, 20%, and 50% labeled data | mIoU: 45.4 -> 51.7; 56.1 -> 59.8; 57.8 -> 61.8; 58.7 -> 60.4 | Helped |
| Add patch-based mixing (ablation) | Cylinder3D semi-supervised baseline · SemanticKITTI, 1%, 10%, 20%, and 50% labeled data | mIoU: 45.4 -> 51.3; 56.1 -> 61.4; 57.8 -> 60.7; 58.7 -> 60.5 | Helped |
| Add instance filling (ablation) | Cylinder3D semi-supervised baseline · SemanticKITTI, 1%, 10%, 20%, and 50% labeled data | mIoU: 45.4 -> 52.3; 56.1 -> 61.7; 57.8 -> 60.1; 58.7 -> 61.1 | Helped |
| Add point erasure, patch-based mixing, and instance filling (ablation) | Cylinder3D semi-supervised baseline · SemanticKITTI, 1%, 10%, and 20% labeled data | mIoU: 45.4 -> 54.5; 56.1 -> 63.3; 57.8 -> 63.7 | Helped |
| Use multi-scene rather than dual-scene interactions in patch augmentation (ablation) | AIScene patch-based data augmentation · SemanticKITTI, average across labeled-data splits | mIoU: 59.7 -> 61.5 (+1.8) | Helped |
Surprising findings
- Multi-scene patch mixing outperforms dual-scene mixing by 1.8 average mIoU on SemanticKITTI.
- The point erasure strategy is plug-and-play and also improves LaserMix when added to it.
Limitations
- Relies on a confidence threshold (0.9) for pseudo-label generation and erasure, tuned experimentally.
- Patch split number n and consistency loss weights are hyperparameters requiring tuning per dataset.
Experimental setup
Datasets: SemanticKITTI nuScenes
Benchmarks: SemanticKITTI nuScenes
Metrics: mIoU
Baselines: Mean Teacher CBST CPS LaserMix GPC IGNet LiM3D LASS3D DDSemi
Builds on: Mean Teacher LaserMix DDSemi Cylinder3D
Closest papers by meaning
- Weakly Supervised Semantic Segmentation for Driving Scenes (AAAI 2024 · similarity 0.88)
- VLM2Scene: Self-Supervised Image-Text-LiDAR Learning with Foundation Models for Autonomous Driving Scene Understanding (AAAI 2024 · similarity 0.88)
- TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on the Line of Sight (NeurIPS 2024 · similarity 0.87)
- Construct to Associate: Cooperative Context Learning for Domain Adaptive Point Cloud Segmentation (CVPR 2024 · similarity 0.87)
- HiLoTs: High-Low Temporal Sensitive Representation Learning for Semi-Supervised LiDAR Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.87)
- Beyond the Label Itself: Latent Labels Enhance Semi-supervised Point Cloud Panoptic Segmentation (AAAI 2024 · 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.