Research map / Point clouds / Segmentation and detection
Data Efficient 3D Learner via Knowledge Transferred from 2D Model
Ping-Chung Yu, Cheng Sun, Min Sun
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 |
|---|---|---|---|
| add pre-training with 2D-parser pseudo-labels (panoramic) (ablation) | SparseConv · ScanNet val LA 20 points | mIoU: 51.6 -> 58.5 | Improved |
| add pre-training with 2D-parser pseudo-labels (panoramic) (ablation) | O-CNN · ScanNet val LA 200 points | mIoU: 65.0 -> 67.8 | Improved |
| use panoramic instead of perspective pseudo-label pre-training (ablation) | SparseConv · ScanNet val LA 20 points | mIoU: 56.2 -> 58.5 | Improved |
| use panoramic instead of perspective pseudo-label pre-training (ablation) | O-CNN · ScanNet val LA 20 points | mIoU: 56.2 -> 55.6 | Hurt |
| use ground-truth labels instead of pseudo-labels for perspective pre-training (ablation) | SparseConv · ScanNet val LA 20 points | mIoU: 56.2 -> 55.4 | No clear effect |
| use ground-truth labels instead of pseudo-labels for perspective pre-training (ablation) | O-CNN · ScanNet val LA 20 points | mIoU: 56.2 -> 56.9 | No clear effect |
| add semi-supervised learning on top of pre-training (ablation) | Our pre-training · ScanNet val LA 20 points | mIoU: 55.9 -> 61.1 | Improved |
| add semi-supervised learning to model trained from scratch (ablation) | Trained from scratch · ScanNet val LA 20 points | mIoU: 54.2 -> 56.6 | Improved |
| add semi-supervised learning to pre-trained model | Our pre-training · ScanNet LR test 1% | mIoU: 26.6 -> 26.3 | No clear effect |
| add semi-supervised learning to pre-trained model | Our pre-training · ScanNet LR test 5% | mIoU: 46.7 -> 50.8 | Improved |
| add pre-training with semi-supervised learning (ablation) | O-CNN · ScanNet val LR 5% | mIoU: 44.9 -> 49.4 | Improved |
| add pre-training to object classification training | O-CNN · ModelNet40 1% | accuracy: 60.2 -> 65.5 | Improved |
Closest papers by meaning
- SSR-2D: Semantic 3D Scene Reconstruction From 2D Images (TPAMI 2024 · similarity 0.89)
- Geometry and Uncertainty-Aware 3D Point Cloud Class-Incremental Semantic Segmentation (CVPR 2023 · similarity 0.89)
- Bayesian Self-Training for Semi-Supervised 3D Segmentation (ECCV 2024 · similarity 0.89)
- 3D Segmenter: 3D Transformer based Semantic Segmentation via 2D Panoramic Distillation (ICLR 2023 · similarity 0.88)
- Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts (CVPR 2021 · similarity 0.88)
- DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation (ICLR 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.
