Research map / Label-efficient and robust learning / Semi-supervised learning
Instance Consistency Regularization for Semi-Supervised 3D Instance Segmentation
Yizheng Wu, Zhiyu Pan, Kewei Wang, Xingyi Li, Jiahao Cui, Liwen Xiao, Guosheng Lin, Zhiguo Cao
Read the paperCheck 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 |
|---|---|---|---|
| use instance consistency only instead of joint consistency (ablation) | InsTeacher3D · ScanNetV2 val (20% labels) | mAP: 41.8 -> 42.7 | Improved |
| add semantic pseudo labels on top of instance consistency (semantic only vs instance only) (ablation) | InsTeacher3D · ScanNetV2 val (20% labels) | mAP: 42.7 -> 41.1 | Hurt |
| add mean teacher to the TWIST-style self-training (ablation) | serial model · ScanNetV2 val (20% labels) | mAP: 32.8 -> 39.0 | Improved |
| replace serial model with parallel DKNet (ablation) | JCR with mean teacher · ScanNetV2 val (20% labels) | mAP: 39.0 -> 41.8 | Improved |
| replace sem2ins serial model with ins2sem serial model (ablation) | serial JCR · ScanNetV2 val (20% labels) | mAP: 39.0 -> 34.7 | Hurt |
| add superpoint refinement to dynamic mask generation (ablation) | InsTeacher3D · ScanNetV2 val (20% labels) | mAP: 37.9 -> 41.4 | Improved |
| add instance self-enhancement to dynamic mask generation (ablation) | InsTeacher3D · ScanNetV2 val (20% labels) | AP50: 55.6 -> 57.0 | Improved |
| add both self-enhancement and superpoint refinement (ablation) | InsTeacher3D · ScanNetV2 val (20% labels) | mAP: 37.9 -> 42.7 | Improved |
| use EMA alpha 0.999 instead of 0.99 (ablation) | mean teacher · ScanNetV2 val (20% labels) | mAP: 39.4 -> 42.7 | Improved |
| use EMA alpha 0.9999 instead of 0.999 (ablation) | mean teacher · ScanNetV2 val (20% labels) | mAP: 42.7 -> 42.0 | No clear effect |
| choose centroids by heatmap score instead of randomly (ablation) | DKNet · ScanNetV2 val (20% labels) | mAP: 39.5 -> 42.7 | Improved |
| use centroid distance matching instead of IoU matching (ablation) | DKNet · ScanNetV2 val (20% labels) | mAP: 40.9 -> 42.7 | Improved |
Closest papers by meaning
- TWIST: Two-Way Inter-Label Self-Training for Semi-Supervised 3D Instance Segmentation (CVPR 2022 · similarity 0.91)
- Bayesian Self-Training for Semi-Supervised 3D Segmentation (ECCV 2024 · similarity 0.91)
- Insightful Instance Features for 3D Instance Segmentation (CVPR 2025 · similarity 0.89)
- Sketchy Bounding-box Supervision for 3D Instance Segmentation (CVPR 2025 · similarity 0.89)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · similarity 0.88)
- BSNet: Box-Supervised Simulation-assisted Mean Teacher for 3D Instance Segmentation (CVPR 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.
