Research map / Point clouds / Segmentation and detection
One Thing One Click: A Self-Training Approach for Weakly Supervised 3D Semantic Segmentation
Zhengzhe Liu, Xiaojuan Qi, Chi-Wing Fu
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply iterative self-training to initial pseudo labels (ablation) | 3D U-Net on ScanNet-v2 val · ScanNet-v2 val | mIoU: 60.14 -> 65.91 | Improved |
| Add graph propagation with color and coordinate similarity (ablation) | Self-trained 3D U-Net on ScanNet-v2 val · ScanNet-v2 val | mIoU: 65.91 -> 67.92 | Improved |
| Add a relation network to graph similarity and use the memory bank for category prototypes (ablation) | 3D U-Net+GP on ScanNet-v2 val · ScanNet-v2 val | mIoU: 67.92 -> 70.45 | Improved |
| Use current-minibatch prototypes instead of a memory bank (ablation) | 3D U-Net+Rel+GP on ScanNet-v2 val · ScanNet-v2 val | mIoU: 67.98 -> 70.45 | Hurt |
| Use graph propagation and the relation network during inference (ablation) | 3D U-Net+Rel+GP on ScanNet-v2 val · ScanNet-v2 val | mIoU: 69.12 -> 70.45 | Improved |
| Use self-training with one-point-per-object supervision instead of initial pseudo labels only (ablation) | One Thing One Click on S3DIS Area 5 · S3DIS Area 5 | mIoU: 43.7 -> 50.1 | Improved |
| Use self-training with three points per object instead of initial pseudo labels only (ablation) | One Thing Three Clicks on S3DIS Area 5 · S3DIS Area 5 | mIoU: 48.9 -> 55.3 | Improved |
| Annotate three points rather than one point per object (ablation) | One Thing Click annotation scheme on S3DIS Area 5 · S3DIS Area 5 | mIoU: 50.1 -> 55.3 | Improved |
| Use only the initial pseudo labels instead of iterative self-training (ablation) | One Thing One Click on ScanNet-v2 val · ScanNet-v2 val | mIoU: 62.18 -> 70.45 | Hurt |
| Disable graph propagation and relation network during inference (ablation) | One Thing One Click on ScanNet-v2 val · ScanNet-v2 val | mIoU: 70.45 -> 68.96 | Hurt |
| Disable graph propagation and relation network during inference (ablation) | One Thing One Click on S3DIS Area 5 · S3DIS Area 5 | mIoU: 50.1 -> 49.4 | Hurt |
| Disable graph propagation and relation network during inference (ablation) | One Thing Three Clicks on S3DIS Area 5 · S3DIS Area 5 | mIoU: 55.3 -> 54.1 | Hurt |
Closest papers by meaning
- Quantity-Quality Enhanced Self-Training Network for Weakly Supervised Point Cloud Semantic Segmentation (TPAMI 2025 · similarity 0.92)
- Self-Prediction for Joint Instance and Semantic Segmentation of Point Clouds (ECCV 2020 · similarity 0.91)
- Multi-Path Region Mining for Weakly Supervised 3D Semantic Segmentation on Point Clouds (CVPR 2020 · similarity 0.91)
- Less Is More: Label Recommendation for Weakly Supervised Point Cloud Semantic Segmentation (AAAI 2024 · similarity 0.89)
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 2024 · similarity 0.89)
- Multi-Modality Affinity Inference for Weakly Supervised 3D Semantic Segmentation (AAAI 2024 · similarity 0.88)
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.
