Research map / Point clouds / Shape and surface completion
Self-Supervised Large Scale Point Cloud Completion for Archaeological Site Restoration
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.
Projects large archaeological point clouds into 5-channel multiple-center-of-projection images and performs patch-based self-supervised inpainting (LaMa-derived inpainter with adversarial, consistency, rotation, and texture-similarity losses) to complete missing walls.
Problem
Self-supervised completion of large-scale, colored point clouds of archaeological structures with missing surfaces, no ground truth, and highly unbalanced point distributions (points mostly missing at the top of walls).
Approach
Represent point clouds as MCOP images with RGB, depth and rotation channels so completion becomes image inpainting; use patch-based adversarial self-supervision (fixed-size windows with high completeness) instead of whole-image masking to avoid the completer filling unbalanced regions with default values, plus consistency, rotation-sum, and Gram-matrix texture terms.
Key results
- On Mawchu Incomplete image-level metrics, Ours achieves PSNR 14.28 (texture) and 17.87 (geometry) and FID 90.81 (texture) vs. 203.85 for MCOP+SSII.
- Point-level Chamfer distance of 8.22 cm vs. 9.18 cm for MCOP+SSII(Patch) and 23.57 cm for P2C, best among compared methods.
- Method completes 600+ incomplete archaeological structures in Peru, scaling to point clouds of up to millions of points (e.g., 15 million points).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed completion method instead of MCOP+SSII(Patch) | MCOP point cloud completion · Mawchu Incomplete | CD (cm)↓: 9.18 -> 8.22 | Helped |
| Use the proposed completion method instead of MCOP+SSII | MCOP point cloud completion · Mawchu Incomplete masked regions | MAE geometry↓: 0.143 -> 0.136 | Helped |
| Use the proposed completion method instead of MCOP+SSII(Patch) | MCOP point cloud completion · Mawchu Incomplete masked regions | FID texture↓: 179.05 -> 90.81 | Helped |
| Add patch-based self-supervision rather than applying image-level self-supervision directly (ablation) | MCOP image completion · Unbalanced MCOP images | completion sharpness | Helped |
| Add the texture similarity term (ablation) | MCOP image completion · Completed MCOP images | pattern consistency | Helped |
Surprising findings
- Standard whole-image adversarial inpainting (UAIR) lets the completer escape the penalty by filling unbalanced upper regions with default values.
Limitations
- Relies on prior manual annotation of object boundaries and shadow regions.
- Quantitative evaluation uses synthetically removed points since no ground truth exists.
- Shadow elimination remains imperfect.
Experimental setup
Datasets: Mawchu Raw Huaman Raw Kuelap Raw Mawchu Incomplete
Benchmarks: Mawchu Incomplete
Metrics: MAE SSIM PSNR FID Chamfer distance precision coverage
Baselines: ACL-SPC P2C MCOP+UAIR MCOP+SSII MCOP+SSII(Patch)
Builds on: LaMa UAIR SSII PointPnCNet ACL-SPC
Closest papers by meaning
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.86)
- ComPC: Completing a 3D Point Cloud with 2D Diffusion Priors (ICLR 2025 · similarity 0.85)
- Position-Aware Guided Point Cloud Completion with CLIP Model (AAAI 2025 · similarity 0.85)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.85)
- CRA-PCN: Point Cloud Completion with Intra- and Inter-level Cross-Resolution Transformers (AAAI 2024 · similarity 0.85)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.85)
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.