Research map / Point clouds / Segmentation and detection
Density-Guided Semi-Supervised 3D Semantic Segmentation with Dual-Space Hardness Sampling
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 DDSemi, a semi-supervised 3D semantic segmentation method using density-guided point-to-anchor contrastive learning and dual-space (geometric + feature) hardness sampling.
Problem
Semi-supervised 3D semantic segmentation of large point clouds with few dense labels; existing point-to-point contrastive learning is sensitive to outliers and confusing points, leading to insufficient point-wise representation modeling.
Approach
Estimates a per-class anchor from a category-wise memory bank using the top-density features (cluster centers lie in dense regions), performs point-to-anchor inter- and intra-contrast losses across two perturbed streams, and reweights intra-contrast loss by geometric and feature-space density so hard points in sparse regions get more weight.
Key results
- On SemanticKITTI with 1% labels DDSemi reaches 59.3 mIoU vs 58.4 for LiM3D (previous best); with 10% it reaches 65.1 mIoU.
- On nuScenes with 1% labels DDSemi reaches 58.1 mIoU, and 76.5 at 50% labels, outperforming LaserMix (voxel) 73.2 at 50%.
- On S3DIS with 5% labels DDSemi reaches 63.2 mIoU vs 53.0 for GPC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DDSemi rather than the strongest reported baseline | DDSemi · SemanticKITTI, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 59.3 vs 58.4; 65.1 vs 62.2; 66.3 vs 63.1; 67.0 vs 63.6 | Helped |
| use DDSemi rather than the strongest reported baseline | DDSemi · nuScenes, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 58.1 vs 55.3; 70.2 vs 69.9; 74.0 vs 71.8; 76.5 vs 73.2 | Helped |
| use DDSemi rather than the strongest reported baseline | DDSemi · S3DIS, labeled ratios 5%, 10%, 20%, 30%, and 40% | mIoU (%): 63.2 vs 53.0; 66.8 vs 57.7; 70.3 vs 63.5; 70.5 vs 64.9; 70.8 vs 65.0 | Helped |
| add inter-contrast loss to labeled cross-entropy training (ablation) | DDSemi · SemanticKITTI, 10% labeled | mIoU (%): 56.1 -> 59.7 | Helped |
| add intra-contrast loss to inter-contrast training (ablation) | DDSemi · SemanticKITTI, 10% labeled | mIoU (%): 59.7 -> 61.3 | Helped |
| add DHS reweighting in geometric space (ablation) | DDSemi · SemanticKITTI, 10% labeled | mIoU (%): 61.3 -> 61.9 | Helped |
| add DHS reweighting in feature space (ablation) | DDSemi · SemanticKITTI, 10% labeled | mIoU (%): 61.3 -> 63.8 | Helped |
| add DHS in both geometric and feature spaces (ablation) | DDSemi · SemanticKITTI, 10% labeled | mIoU (%): 63.8 -> 65.1 | Helped |
| replace point-to-point contrast with point-to-anchor contrast (ablation) | DDSemi · SemanticKITTI, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 55.6 -> 59.3; 63.1 -> 65.1; 64.0 -> 66.3; 64.8 -> 67.0 | Helped |
| use mixed progressive and FIFO memory-bank updates instead of progressive-only updates (ablation) | DDSemi · SemanticKITTI, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 58.7 -> 59.3; 64.3 -> 65.1; 65.5 -> 66.3; 66.3 -> 67.0 | Helped |
| use mixed progressive and FIFO memory-bank updates instead of FIFO-only updates (ablation) | DDSemi · SemanticKITTI, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 58.8 -> 59.3; 64.5 -> 65.1; 65.7 -> 66.3; 66.2 -> 67.0 | Helped |
| replace random anchor estimation with density-based estimation (ablation) | DDSemi · SemanticKITTI, labeled ratios 1%, 10%, 20%, and 50% | mIoU (%): 58.2 -> 59.3; 64.0 -> 65.1; 64.9 -> 66.3; 65.8 -> 67.0 | Helped |
Limitations
- Larger memory bank and k increase storage and density computation cost
- Hyperparameters (gamma, tau, alpha, epsilon, lambda weights) are manually set and tuned
Experimental setup
Datasets: SemanticKITTI nuScenes S3DIS
Benchmarks: SemanticKITTI nuScenes S3DIS
Metrics: mIoU
Baselines: MeanTeacher CBST CPS LaserMix GPC LiM3D SSS-Net
Builds on: LaserMix GPC Mean Teacher CPS
Closest papers by meaning
- Hunting Sparsity: Density-Guided Contrastive Learning for Semi-Supervised Semantic Segmentation (CVPR 2023 · similarity 0.88)
- An End-to-End Robust Point Cloud Semantic Segmentation Network with Single-Step Conditional Diffusion Models (CVPR 2025 · similarity 0.88)
- Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation (AAAI 2024 · similarity 0.88)
- SSPC-Net: Semi-supervised Semantic 3D Point Cloud Segmentation Network (AAAI 2021 · similarity 0.88)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.87)
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 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.