Research map / Point clouds / Segmentation and detection
Shape2Scene: 3D Scene Representation Learning Through Pre-training on Shape Data
Tuo Feng, Wenguan Wang, Ruijie Quan, Yi Yang
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 |
|---|---|---|---|
| pre-train MH-P with S2SS and PPC instead of training from scratch (ablation) | MH-P · ScanObjectNN PB_T50_RS | overall accuracy (%): 85.9 -> 87.8 | Improved |
| pre-train MH-V with S2SS and PPC instead of training from scratch (ablation) | MH-V · S3DIS Area5 | mIoU (%): 70.3 -> 72.7 | Improved |
| replace single-shape pre-training with S2SS pseudo scenes (ablation) | MH-V · S3DIS Area5 | mIoU (%): 70.5 -> 72.7 | Improved |
| pre-train on single shape data without pseudo scenes (ablation) | MH-V · S3DIS Area5 | mIoU (%): 70.3 -> 70.5 | No clear effect |
| scale pseudo-scene pre-training data to UHD and LHD instead of ScanNet real scenes (ablation) | MH-V · S3DIS Area5 | mIoU (%): 73.5 -> 74.1 | Improved |
| use four scales instead of two (ablation) | MH-P · ScanObjectNN PB_T50_RS | overall accuracy (%): 82.3 -> 87.8 | Improved |
| use four scales instead of two (ablation) | MH-V · S3DIS Area5 | mIoU (%): 62.3 -> 72.7 | Improved |
| integrate high-resolution features from all scales into final output instead of only the last scale (ablation) | MH-V · S3DIS Area5 | mIoU (%): 71.6 -> 72.7 | Improved |
| increase shapes per pseudo scene M beyond the optimum (ablation) | S2SS · S3DIS Area5 | mIoU (%): 72.7 (M=6) vs 72.5 (M=8), 72.6 (M=10) | No clear effect |
| increase M from 1 to 4 shapes per scene (ablation) | S2SS · ScanObjectNN PB_T50_RS | overall accuracy (%): 86.8 -> 87.8 | Improved |
| use MH-P backbone instead of PointGPT-S transformer, ShapeNet pre-training | MH-PS · ModelNet40 | overall accuracy (%): 94.0 -> 94.1 | No clear effect |
| pre-train MH-V on pseudo scenes for outdoor detection and segmentation versus ShapeNet-pretrained PointDif | MH-VS · SemanticKITTI val | mIoU (%): 71.3 -> 70.8 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Point-M2AE — Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.90)
- PointGLR: Unsupervised Structural Representation Learning of 3D Point Clouds (TPAMI 2022 · similarity 0.89)
- 3D-RCNN: Instance-Level 3D Object Reconstruction via Render-and-Compare (CVPR 2018 · similarity 0.89)
- PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding (ECCV 2020 · similarity 0.89)
- CLIP2Scene: Towards Label-Efficient 3D Scene Understanding by CLIP (CVPR 2023 · similarity 0.89)
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · 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.
