Research map / Point clouds / Segmentation and detection
Exploring Data-Efficient 3D Scene Understanding With Contrastive Scene Contexts
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 spatial scene-context partitions to point-level contrastive pre-training | PointContrast-style pre-training · S3DIS instance segmentation | mAP@0.5: 60.5 -> 63.4 | Helped |
| Add spatial scene-context partitions to pre-training | ScanNet-LA, 200 labeled points per scene · ScanNet-LA instance segmentation | mAP@0.5: 44.5 -> 48.9 | Helped |
| Initialize VoteNet with Contrastive Scene Contexts pre-trained weights | VoteNet · ScanNet-LA, 7 labeled boxes per scene | mAP@0.5: 26.5 -> 30.4 | Helped |
| Initialize segmentation with Contrastive Scene Contexts pre-trained weights | Sparse Res-UNet · ScanNet-LR, 1% of training scenes | mAP@0.5: 9.9 -> 13.2 | Helped |
| Initialize VoteNet with Contrastive Scene Contexts pre-trained weights | VoteNet · ScanNet-LR, 20% of training scenes | mAP@0.5: 4.6 -> 20.9 | Helped |
| Use Contrastive Scene Contexts instead of PointContrast pre-training | Sparse Res-UNet · ScanNet validation semantic segmentation | mIoU: 74.1 -> 73.8 | Hurt |
| Increase sampled points without using spatial partitions (ablation) | PointContrast-style pre-training · S3DIS instance segmentation | mAP@0.5 | No clear effect |
| Add scene-context partitions to enable using more sampled points (ablation) | Contrastive Scene Contexts pre-training · S3DIS instance segmentation | mAP@0.5 | Helped |
| Use 8 spatial partitions (ablation) | Contrastive Scene Contexts pre-training · S3DIS instance segmentation | mAP@0.5 | Helped |
| Select annotation points by clustering pre-trained features rather than random sampling or raw-input k-means | Active labeling · ScanNet-LA, 20 labeled points | mAP@0.5 and mIoU: >10% absolute improvement | Helped |
| Use active labeling together with pre-trained initialization | ScanNet-LA segmentation · ScanNet-LA, limited point annotations | segmentation performance: 1% boost | Helped |
| Use pre-trained feature-space clustering rather than raw RGB+XYZ k-means sampling | Active labeling in cluttered ScanNet scenes · 200 cluttered scenes | distinct-object coverage | Helped |
Closest papers by meaning
- CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP (CVPR 2023 · similarity 0.90)
- CLIP2: Contrastive Language-Image-Point Pretraining From Real-World Point Cloud Data (CVPR 2023 · similarity 0.90)
- Relation3D : Enhancing Relation Modeling for Point Cloud Instance Segmentation (CVPR 2025 · similarity 0.89)
- RegionPLC: Regional Point-Language Contrastive Learning for Open-World 3D Scene Understanding (CVPR 2024 · similarity 0.89)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.88)
- Mitigating Object Dependencies: Improving Point Cloud Self-Supervised Learning through Object Exchange (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.