AnchorComplete: Recovering Complete Scene Meshes from Partial Point Clouds
Organizations: Technical University of Munich · Munich Center for Machine Learning · Nanyang Technological University · University of Oxford
Abstract
Most image-based 3D reconstructors primarily recover visible surfaces, producing partial point clouds with large holes and leaving unobserved geometry unresolved. To address this limitation, we introduce AnchorComplete, a framework that reconstructs complete scene meshes from partial point clouds while preserving observed geometry and supporting scenes of varying scales. Building on TRELLIS.2, our method combines partial-point conditioning, which integrates global scene tokens with local spatial evidence to anchor generation to the input, and metric-scale coarse-to-fine generation, which progressively refines scene geometry within fixed-size windows while maintaining global structural consistency. Experiments on Replica and SCRREAM demonstrate that AnchorComplete consistently outperforms strong baselines in complete-scene reconstruction, particularly under sparse-view settings.
Figures & tables
| Method | Replica | SCRREAM | Runtime | ||||||
|---|---|---|---|---|---|---|---|---|---|
| F 0.1 | F 0.05 | F 0.1 | F 0.05 | ||||||
| Feed-forward dense reconstruction | |||||||||
| VGGT- ( Wang et al., 2026 ) | 0.062 | 0.230 | 0.560 | 0.412 | 0.062 | 0.247 | 0.582 | 0.433 | <1s |
| DA3 ( Lin et al., 2025 ) | 0.048 | 0.220 | 0.611 | 0.508 | 0.052 | 0.227 | 0.607 | 0.453 | <1s |
| CUT3R ( Wang et al., 2025b ) | 0.244 | 0.360 | 0.355 | 0.214 | 0.189 | 0.359 | 0.455 | 0.311 | <1s |
| PI-3 ( Wang et al., 2025c ) | 0.053 | 0.224 | 0.573 | 0.426 | 0.096 | 0.253 | 0.485 | 0.315 | <1s |
| (a) Partial-point conditioning | |||||||
|---|---|---|---|---|---|---|---|
| Method | Global | Local | F 0.1 | F 0.05 | F 0.02 | ||
| TRELLIS ∗ | ✗ | ✗ | 0.244 | 0.288 | 0.308 | 0.150 | 0.030 |
| w/o local injection | ✓ | ✗ | 0.072 | 0.105 | 0.749 | 0.532 | 0.131 |
| w/o global tokens | ✗ | ✓ | 0.060 | 0.086 | 0.787 | 0.652 | 0.219 |
| Ours | ✓ | ✓ | 0.055 | 0.067 | 0.819 | 0.684 | 0.252 |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | Input source | F 0.1 | F 0.05 | ||
|---|---|---|---|---|---|
| Per-view point map | DA3 (default) | 0.055 | 0.067 | 0.819 | 0.684 |
| Pi3 | 0.048 | 0.062 | 0.837 | 0.655 | |
| VGGT- | 0.068 | 0.079 | 0.746 | 0.547 | |
| SfM/MVS | COLMAP | 0.069 | 0.084 | 0.770 | 0.641 |
| Visual SLAM | DROID-SLAM | 0.044 | 0.053 | 0.870 | 0.759 |
| DPVO | 0.046 | 0.055 | 0.872 | 0.713 |
3D Scene Completion
3D scene completion infers missing geometry, occupancy, semantics, and appearance from partial images, point clouds, or depth scans.
See all 44 papers