Research map / Point clouds / Segmentation and detection
Distribution Guidance Network 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 DGNet, which aligns the embedding distribution of weakly supervised point cloud segmentation networks to a mixture of von Mises-Fisher distributions via a nested EM procedure, improving results on S3DIS, ScanNetV2 and SemanticKITTI.
Problem
Weakly supervised point cloud semantic segmentation suffers from inadequate supervision with sparse labels; existing constraints are heuristic and ignore the distribution of network embeddings.
Approach
Identify moVMF with cosine similarity as the best descriptor of the semantic feature space, then dynamically align network embeddings to it with a nested soft-moVMF EM procedure, combined with truncated cross-entropy, a vMF likelihood loss, a discriminative loss and a consistency loss.
Key results
- On S3DIS Area 5 at 0.01% labels, PointNeXt+DGNet reaches 62.4 mIoU vs 58.4 for PointNeXt baseline.
- On S3DIS at 0.1% labels, PointNeXt+DGNet reaches 67.8 mIoU, about 97% of the fully supervised PointNeXt (69.2).
- On SemanticKITTI, DGNet (RandLA-Net) outperforms SQN by 1.0% and 2.1% mIoU at 0.1% and 0.01% label rates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DGNet distribution guidance to PointNeXt | PointNeXt · S3DIS Area 5, 0.1% | mIoU: 65.0 -> 67.8 | Helped |
| Add DGNet distribution guidance to PointNeXt | PointNeXt · S3DIS Area 5, 0.01% | mIoU: 58.4 -> 62.4 | Helped |
| Add DGNet distribution guidance | RandLA-Net · SemanticKITTI, 0.1% | mIoU: +1.0 mIoU | Helped |
| Add DGNet distribution guidance | RandLA-Net · SemanticKITTI, 0.01% | mIoU: +2.1 mIoU | Helped |
| Use moVMF with cosine similarity instead of category prototype or GMM alternatives (ablation) | DGNet distribution alignment · S3DIS Area 5, distribution selection | mIoU: 59.9; 60.3; 61.3 -> 62.4 | Helped |
| Replace partial cross-entropy with truncated cross-entropy (ablation) | DGNet · S3DIS Area 5, loss terms | mIoU: 58.4 -> 59.1 | Helped |
| Add soft-assignment vMF loss (ablation) | DGNet with truncated cross-entropy · S3DIS Area 5, loss terms | mIoU: 59.1 -> 60.4 | Helped |
| Add discriminative loss (ablation) | DGNet with truncated cross-entropy · S3DIS Area 5, loss terms | mIoU: 59.1 -> 59.8 | Helped |
| Add consistency loss (ablation) | DGNet with truncated cross-entropy · S3DIS Area 5, loss terms | mIoU: 59.1 -> 61.1 | Helped |
| Replace soft-assignment vMF loss with hard assignment (ablation) | DGNet · S3DIS Area 5, loss terms | mIoU: 60.4 -> 58.9 | Hurt |
| Update both α and μ using soft-moVMF rather than partial or alternative updates (ablation) | DGNet Nested EM · S3DIS Area 5, Nested EM | mIoU: 61.0; 60.9; 60.0; 59.8; 60.2 -> 62.4 | Helped |
| Use soft assignment instead of hard-moVMF or KNN-moVMF (ablation) | DGNet Nested EM · S3DIS Area 5, Nested EM | mIoU: 2.6% and 2.2% mIoU improvements | Helped |
Surprising findings
- Under the ablation, a single hard-assignment moVMF variant or simply optimizing only some parameters degrades performance relative to the full soft-moVMF update.
- The DGNet gain on ScanNetV2 is smaller than on S3DIS because the cross-entropy term already gets pseudo-label support there.
Limitations
- Exploration of embedding distribution is described by the authors as preliminary.
- More sophisticated feature distributions may overfit under sparse annotations.
- No error bars reported, following other weak supervision methods.
Experimental setup
Datasets: S3DIS ScanNetV2 SemanticKITTI
Benchmarks: S3DIS Area 5 ScanNetV2 validation SemanticKITTI sequence 08
Metrics: mIoU
Baselines: PointNet SQN HybridCR ERDA DeepGCN PointNeXt PointTransV1 CPCM PointMatch AADNet PSD DCL
Builds on: moVMF soft-moVMF EM PointNeXt DeepGCN LaDS
Closest papers by meaning
- Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations (AAAI 2025 · similarity 0.88)
- Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation (AAAI 2024 · similarity 0.88)
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 2024 · similarity 0.88)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.87)
- DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation (ICML 2025 · similarity 0.85)
- Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation (AAAI 2025 · 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.