Research map / Point clouds / Segmentation and detection
Less Is More: Label Recommendation for Weakly Supervised Point Cloud 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 a three-stage label recommendation framework (upsampling-based inductive bias learning, dual-tier cross-scene clustering to recommend points to label, and LabelAttention-based weakly supervised segmentation) for point cloud semantic segmentation under extremely sparse labels.
Problem
Weakly supervised point cloud semantic segmentation relies on sparse annotations whose form (random or one-thing-one-click) is unconstrained or heuristic, yielding suboptimal label efficiency and supervisory information deficiency.
Approach
Rather than pre-training or active learning, learn inductive bias via 2x point cloud upsampling (a generative pretext task chosen to minimize distribution shift), then use dual-tier cross-scene K-means cluster centers as recommended points for annotators, and train the segmenter with a LabelAttention module that directly links labeled points plus position encoding.
Key results
- At 0.01% label rate on ScanNetV2, the method reaches 67.4 mIoU at 1% label rate and 63.7 mIoU with 20 pts/scene (super-voxel), surpassing AADNet's 62.5.
- On S3DIS Area 5 at 0.01% label rate, the method achieves 62.9 mIoU versus 60.8 for AADNet and 59.9 for PointMatch, and improves 4.5% over random sampling baseline at 0.01%.
- Pretext ablation: 2x point upsampling gives 49.6 mIoU vs 43.7 for Hardest-Contrastive and 45.1 for point colorize; LabelAttention gives 56.1 mIoU vs 55.8 for GlobalAttention.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use label recommendation instead of competing weakly supervised methods | Ours · ScanNetV2 test, 1% supervision | mIoU: 66.8 -> 67.4 | Helped |
| use label recommendation instead of AADNet | Ours · ScanNetV2 test, 20 pts/scene | mIoU: 62.5 -> 63.7 | Helped |
| use label recommendation instead of AADNet | Ours · S3DIS Area 5, 0.1% supervision | mIoU: 67.2 -> 68.9 | Helped |
| use label recommendation instead of AADNet | Ours · S3DIS Area 5, 0.01% supervision | mIoU: 60.8 -> 62.9 | Helped |
| use 2× point upsampling rather than contrastive or alternative generative pretext tasks (ablation) | Label recommendation framework · S3DIS Area 5 | mIoU: 43.7; 42.6; 45.1; 49.6; 48.2 | Helped |
| increase point upsampling rate from 2× to 4× (ablation) | Label recommendation framework · S3DIS Area 5 | mIoU: 49.6 -> 42.9 | Hurt |
| increase point upsampling rate from 2× to 8× (ablation) | Label recommendation framework · S3DIS Area 5 | mIoU: 49.6 -> 43.2 | Hurt |
| vary cross-scene clustering batch size (ablation) | Label recommendation framework · S3DIS Area 5 | mIoU: 49.6; 50.3; 51.2; 48.6; 47.8 | Mixed |
| add position encoding (ablation) | PointNeXt-L · S3DIS Area 5 | mIoU: 51.2 -> 55.0 | Helped |
| add random attention with position encoding (ablation) | PointNeXt-L · S3DIS Area 5 | mIoU: 55.0 -> 54.4 | Hurt |
| add local attention with position encoding (ablation) | PointNeXt-L · S3DIS Area 5 | mIoU: 55.0 -> 55.7 | Helped |
| add global attention with position encoding (ablation) | PointNeXt-L · S3DIS Area 5 | mIoU: 55.0 -> 55.8 | Helped |
Surprising findings
- Contrastive pretext tasks, which work well for pre-training, underperform generative point-level upsampling within the label recommendation framework.
- Global attention gives similar gains to LabelAttention in ablations, but LabelAttention is cheaper in GFLOPs (30.0 vs 28.7 is actually lower for global; the paper notes LabelAttention is the best overall).
Limitations
- Results depend on the choice of clustering (K-means) and batch size; performance drops for large scene batches.
- Requires a separate inductive bias learning and recommendation stage before segmentation training.
Experimental setup
Datasets: S3DIS ScanNetV2
Benchmarks: S3DIS ScanNetV2
Metrics: mIoU
Baselines: SQN PointMatch OTOC++ AADNet GaIA PSD HybirdCR MILTrans DAT PointNeXt
Builds on: PointNeXt PointNet++ AAD LESS
Closest papers by meaning
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 2024 · similarity 0.92)
- Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations (AAAI 2025 · similarity 0.89)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.89)
- Distribution Guidance Network for Weakly Supervised Point Cloud Semantic Segmentation (NeurIPS 2024 · similarity 0.88)
- Density-Guided Semi-Supervised 3D Semantic Segmentation with Dual-Space Hardness Sampling (CVPR 2024 · similarity 0.88)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic 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.