Research map / Point clouds / Shape and surface completion
Point-based Instance Completion with Scene Constraints
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a point-cloud instance completion model that handles arbitrary object pose/scale and integrates sparse scene constraints (offset shells of the partial scan) via cross-attention, plus a new ScanWCF dataset of watertight, collision-free indoor scenes.
Problem
Point-based object completion methods assume canonically aligned objects and ignore scene constraints (observed surfaces, free space, other objects), while existing instance scene completion methods lag in completion quality and use collision-prone or misaligned ground truth data.
Approach
Predict the object center and seed offsets from it so the model is pose/scale robust, use VI-PointConv encoder with normals, and inject sparse scene constraints (two bounding shells P ± δN of the partial scan representing free and occluded space) into the seed generator through cross-attention.
Key results
- On ScanWCF instance scene completion, ours (w/ pre-training) reaches 47.55 IoU@0.5 mAP vs 8.67 for DIMR and 7.25 for RfD-Net.
- Scene constraints reduce CD from 22.37 to 20.74 and collision COL from 3.75 to 2.67 (about 29% relative improvement in collision distance).
- Ablation: replacing the seed generator with the proposed center-plus-offset design improves seed CD from 47.00 to 42.86 on chairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add VI-PointConv and input normals (ablation) | object completion model · ShapeNet chair completion, arbitrary pose and scale | One-sided CD: 7.12 -> 6.04 | Helped |
| add object-center prediction and predict seeds as offsets (ablation) | object completion model · ShapeNet chair completion, arbitrary pose and scale | CD (Seeds): 47.00 -> 42.86 | Helped |
| add global attention to upsampling layers (ablation) | object completion model · ShapeNet chair completion, arbitrary pose and scale | CD (Dense): 21.11 -> 20.05 | Helped |
| add surface normal prediction module (ablation) | object completion model · ShapeNet chair completion, arbitrary pose and scale | CD (Dense): 20.05 -> 20.08 | No clear effect |
| add scene constraints (ablation) | scene completion model · ScanWCF scene completion | UHD: 60.18 -> 60.07 | No clear effect |
| add scene constraints (ablation) | scene completion model · ScanWCF scene completion | One-sided CD: 12.20 -> 12.10 | No clear effect |
| add scene constraints (ablation) | scene completion model · ScanWCF scene completion | CD: 22.37 -> 20.74 | Helped |
| add scene constraints (ablation) | scene completion model · ScanWCF scene completion | COL: 3.75 -> 2.67 | Helped |
| add pre-training | proposed completion model · ScanWCF instance scene completion | IoU@0.5 mAP: 45.13 -> 47.55 | Helped |
| add pre-training | proposed completion model · ScanWCF scene completion (ground-truth masks) | COL: 3.42 -> 2.67 | Helped |
| use object completion model with pre-training rather than RfD-Net | ScanWCF instance scene completion · ScanWCF instance scene completion | IoU@0.25 mAP: 38.78 -> 62.77 | Helped |
| use proposed model rather than RfD-Net | scene completion with ground-truth instance masks · ScanWCF scene completion | CD: 38.36 -> 20.74 | Helped |
Surprising findings
- Pre-training is not strictly required: the model without pre-training still outperforms RfD-Net and DIMR.
- RfD-Net and DIMR outperform the proposed method on extremely sparse inputs (0-10% of object visible).
Limitations
- Completion framework is deterministic and produces only one completion per partial scan.
- Instance segmentation is not trained jointly with completion; Mask3D is used only at inference.
- Evaluation restricted to 13 categories with more than 150 training examples.
Experimental setup
Datasets: ScanWCF Scan2CAD ScanARCW ScanNet ShapeNet
Benchmarks: ScanWCF
Metrics: IoU mAP Chamfer Distance mAP Light Field Distance mAP Point Coverage Ratio mAP UHD One-sided CD CD COL %COL
Baselines: RfD-Net DIMR SeedFormer SCARP
Builds on: SeedFormer VI-PointConv NKSR Mask3D PointConv
Closest papers by meaning
- Learning Local Displacements for Point Cloud Completion (CVPR 2022 · similarity 0.90)
- Skeleton-bridged Point Completion: From Global Inference to Local Adjustment (NeurIPS 2020 · similarity 0.89)
- Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.88)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.88)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.87)
- SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance (AAAI 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.