Research map / Point clouds / Segmentation and detection
Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations
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.
Analyzes how sparse and inhomogeneous point-cloud annotations affect weakly supervised segmentation via gradient sampling theory, and proposes AADNet (label-aware downsampling plus multiplicative dynamic entropy with asynchronous training) that improves over PointNeXt and RandLA-Net baselines.
Problem
Weakly supervised point cloud semantic segmentation assumes uniformly distributed sparse annotations, but real-world annotations are inhomogeneous, which biases gradient learning and degrades performance.
Approach
Uses a gradient sampling approximation with probability density functions to show that annotation inhomogeneity adds a gradient bias; counters it with a label-aware voxel/point downsampling that boosts labeled-point coverage and a multiplicative entropy calibration weight trained asynchronously with an entropy loss.
Key results
- On S3DIS Area 5 with 0.01% uniform labels, AADNet (PointNeXt) reaches 60.8 mIoU versus 44.7 for the PointNeXt baseline.
- With non-uniform 1% labels on S3DIS (G=1), AADNet (PointNeXt) gets 39.0 mIoU compared with 35.9 for PointNeXt.
- On ScanNetV2 with 20 points per scene, AADNet (PointNeXt) achieves 62.5 mIoU under the supervoxel setup, versus 54.6 for the baseline without supervoxels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply AADNet with RandLA-Net backbone to inhomogeneous annotations | RandLA-Net · S3DIS, 10% label rate, G=1 | mIoU: 43.3 -> 50.5 | Helped |
| Apply AADNet with PointNeXt backbone to inhomogeneous annotations | PointNeXt · S3DIS, 1% label rate, G=1 | mIoU: 35.9 -> 39.0 | Helped |
| Apply AADNet with PointNeXt backbone to uniformly sparse annotations | PointNeXt · S3DIS, 0.01% uniform | mIoU: 44.7 -> 60.8 | Mixed |
| Apply AADNet with PointNeXt backbone to uniformly sparse annotations | PointNeXt · ScanNetV2, 20 points/scene uniform | mIoU: 54.6 -> 62.5 | Helped |
| Apply AADNet with PointNeXt backbone to uniformly sparse annotations | PointNeXt · SemanticKITTI, 0.1% uniform | mIoU: 48.9 -> 53.3 | Helped |
| Add point-level label-first sampling while retaining voxel-level random sampling (ablation) | PointNeXt-L · S3DIS Area 5, 0.01% uniform | mIoU: 44.7 -> 58.4 | Helped |
| Add voxel-level label-first sampling instead of voxel-level random sampling (ablation) | PointNeXt-L · S3DIS Area 5, 0.01% uniform | mIoU: 44.7 -> 45.6 | No clear effect |
| Use label-first sampling at both point and voxel levels (ablation) | PointNeXt-L · S3DIS Area 5, 10% inhomogeneous | mIoU: 54.1 -> 53.8 | Hurt |
| Add multiplicative dynamic entropy with asynchronous training (ablation) | PointNeXt-L · S3DIS Area 5, 0.01% uniform | mIoU: 44.7 -> 49.7 | Helped |
| Add multiplicative dynamic entropy with asynchronous training (ablation) | PointNeXt-L · S3DIS Area 5, 10% inhomogeneous | mIoU: 54.1 -> 56.0 | Helped |
| Replace dynamic entropy calibration with fixed inverse local-density calibration (ablation) | PointNeXt-L · S3DIS Area 5, 0.01% uniform | mIoU: 49.7 -> 43.3 | Hurt |
| Replace multiplicative dynamic entropy with additive dynamic entropy (ablation) | PointNeXt-L · S3DIS Area 5, 0.01% uniform | mIoU: 49.7 -> 48.2 | Hurt |
Surprising findings
- Simultaneous training of entropy loss with partial cross-entropy prevents the partial loss from converging, requiring asynchronous training.
- Even under uniform sparse annotations at extreme sparsity, the distribution becomes inhomogeneous, so the calibration function still helps.
Limitations
- Inhomogeneous annotations are simulated via cluster sampling rather than collected from real annotators.
- Entropy-based calibration is a heuristic alternative to the ideal inverse-density weight, which is unknown in practice.
Experimental setup
Datasets: S3DIS ScanNetV2 SemanticKITTI
Benchmarks: S3DIS Area 5 ScanNetV2 SemanticKITTI validation
Metrics: mIoU
Baselines: SQN EDRA RandLA-Net DCL CPCM PointNeXt RPSC PointMatch MILTrans OTOC
Builds on: PointNeXt RandLA-Net partial cross-entropy loss
Closest papers by meaning
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 2024 · similarity 0.91)
- Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation (AAAI 2024 · similarity 0.89)
- Distribution Guidance Network for Weakly Supervised Point Cloud Semantic Segmentation (NeurIPS 2024 · similarity 0.88)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.86)
- Point Cloud Pre-training with Diffusion Models (CVPR 2024 · similarity 0.86)
- Learning Continuous Implicit Field with Local Distance Indicator for Arbitrary-Scale Point Cloud Upsampling (AAAI 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.