Research map / Point clouds / Shape and surface completion
Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion
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 two-stage point cloud completion method with a diffusion-based coarse generator and a Context-aware Refiner that uses mixed sampling, surface freezing, and rigid-transformation-invariant similarity modeling to exploit short- and long-range contextual information from the partial input.
Problem
High-fidelity texture recovery in the fine stage of point cloud completion, which is ill-posed, given occlusion-induced incompleteness of partial point clouds.
Approach
Refines a diffusion-generated coarse point cloud using contextual information from the partial input: short-range via mixed sampling with surface freezing, and long-range via learned rotation and symmetry-plane alignment that yields rigid-transformation-invariant patch similarity for non-local refinement.
Key results
- On ShapeNet-ViPC, average L2 CD of 2.148 x10^-3 and F-Score@0.001 of 0.719, outperforming CSDN (2.570 CD).
- On MVP, CD of 5.49 x10^-4 is best among compared methods, with F-Score@0.01 of 0.515 versus PDR's 0.499.
- On PartNet, average UHD reduced by 41.4% relative to the next-best method, while the average MMD (1.51) is not the best (Diverse 1.49).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add mixed sampling (ablation) | CRef · ShapeNet-ViPC (ablation slice) | L2 Chamfer Distance × 10^-3 (Avg.): 2.96 -> 2.21 | Helped |
| add surface freezing (ablation) | CRef · ShapeNet-ViPC (ablation slice) | L2 Chamfer Distance × 10^-3 (Avg.): 2.74 -> 2.21 | Helped |
| add rigid transformation (ablation) | CRef · ShapeNet-ViPC (ablation slice) | L2 Chamfer Distance × 10^-3 (Avg.): 2.43 -> 2.21 | Helped |
| replace the full similarity modeling with Euclidean similarity only (ablation) | CRef · ShapeNet-ViPC (ablation slice) | L2 Chamfer Distance × 10^-3 (Avg.): 2.21 -> 2.24 | Hurt |
| use the feature similarity variant instead of the full similarity modeling (ablation) | CRef · ShapeNet-ViPC (ablation slice) | L2 Chamfer Distance × 10^-3 (Avg.): 2.21 -> 2.22 | Hurt |
| use the proposed method instead of CSDN | ShapeNet-ViPC · ShapeNet-ViPC | L2 CD × 10^-3 (Avg.): 2.570 -> 2.148 | Helped |
| use the proposed method instead of ShapeFormer | PartNet · PartNet | MMD × 10^-3 (Avg.): 1.32 -> 1.51 | Hurt |
| use the proposed method instead of PDR | MVP · MVP | EMD × 10^-2: 1.37 -> 2.22 | Hurt |
Surprising findings
- On MVP, the method has the best CD and F-score but not the best EMD (2.22 vs PDR's 1.37).
- Removing the rigid transformation module in the ablation only slightly degrades results (2.43 vs 2.21 average CD on the ablation slice).
Limitations
- Fine stage faces the ill-posed nature of completion, so high-fidelity recovery depends on partial-input redundancy such as symmetry.
- Results on MVP do not win on every metric (EMD and MMD not best).
Experimental setup
Datasets: ShapeNet-ViPC PartNet 3D-EPN MVP
Benchmarks: ShapeNet-ViPC 3D-EPN PartNet MVP
Metrics: L2 Chamfer Distance F-Score EMD MMD UHD
Baselines: PCN TopNet GRNet PF-Net PoinTr ViPC SeedFormer CSDN PointAttN PDR cGAN Diverse
Builds on: DDPM PointTr DGCNN PDR
Closest papers by meaning
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.90)
- PCDreamer: Point Cloud Completion Through Multi-view Diffusion Priors (CVPR 2025 · similarity 0.89)
- PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion (AAAI 2025 · similarity 0.89)
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.87)
- GeoComplete: Geometry-Aware Diffusion for Reference-Driven Image Completion (NeurIPS 2025 · similarity 0.87)
- Position-Aware Guided Point Cloud Completion with CLIP Model (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.