Research map / Detection and segmentation / Weakly supervised detection and segmentation
TWIST: Two-Way Inter-Label Self-Training for Semi-Supervised 3D Instance Segmentation
Ruihang Chu, Xiaoqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, Jiaya Jia
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 |
|---|---|---|---|
| Use TWIST self-training with semantic and instance pseudo labels | 3D instance segmentation · ScanNet v2 validation, 1% labels | mAP: 5.1 -> 9.6 | Improved |
| Initialize with CSC-pretrained weights | TWIST · ScanNet v2 validation, 1% labels | mAP: 9.6 -> 11.5 | Improved |
| Add naive-threshold pseudo semantic labels (ablation) | Supervised-only 3D instance segmentation · ScanNet v2 validation, 5% labels | mAP: 18.2 -> 20.9 | Improved |
| Add naive-threshold pseudo offset vectors (ablation) | Supervised-only 3D instance segmentation · ScanNet v2 validation, 5% labels | mAP: 18.2 -> 19.7 | Improved |
| Use both naive-threshold semantic and offset pseudo labels (ablation) | Supervised-only 3D instance segmentation · ScanNet v2 validation, 5% labels | mAP: 20.9 -> 21.3 | Improved |
| Generate pseudo labels at proposal level (ablation) | TWIST self-training · ScanNet v2 validation, 5% labels | mAP: 21.3 -> 24.8 | Improved |
| Add proposal re-correction for object-level denoising (ablation) | Proposal-based pseudo-label update · ScanNet v2 validation, 5% labels | mAP: 24.8 -> 27.0 | Improved |
| Remove offset-to-semantic interaction and use point-wise thresholding (ablation) | TWIST · ScanNet v2 validation, 5% labels | mAP: 27.0 -> 22.5 | Hurt |
| Remove semantic-to-offset interaction using class-agnostic clustering (ablation) | TWIST · ScanNet v2 validation, 5% labels | mAP: 27.0 -> 21.7 | Hurt |
| Use backbone features instead of RGB features for re-correction (ablation) | Proposal re-correction module · ScanNet v2 validation, 5% labels | mAP: 22.3 -> 25.6 | Improved |
| Add geometric transformation augmentation (ablation) | Proposal re-correction module · ScanNet v2 validation, 5% labels | mAP: 25.6 -> 25.9 | Improved |
| Add geometric mutation augmentation (ablation) | Proposal re-correction module · ScanNet v2 validation, 5% labels | mAP: 25.6 -> 26.6 | Improved |
Closest papers by meaning
- Instance Consistency Regularization for Semi-Supervised 3D Instance Segmentation (TPAMI 2024 · similarity 0.91)
- DBGroup: Dual-Branch Point Grouping for Weakly Supervised 3D Semantic Instance Segmentation (AAAI 2026 · similarity 0.90)
- Bayesian Self-Training for Semi-Supervised 3D Segmentation (ECCV 2024 · similarity 0.89)
- Beyond Semantic to Instance Segmentation: Weakly-Supervised Instance Segmentation via Semantic Knowledge Transfer and Self-Refinement (CVPR 2022 · similarity 0.88)
- Hierarchical Intra-modal Correlation Learning for Label-free 3D Semantic Segmentation (CVPR 2024 · similarity 0.87)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · 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.
