AnchorComplete: Recovering Complete Scene Meshes from Partial Point Clouds

AnchorComplete reconstructs complete scene meshes from partial point clouds by anchoring a TRELLIS.2-based generator to observed points and generating metric-scale coarse-to-fine windows.

Image-based 3D reconstructors primarily recover visible surfaces, so partial point clouds contain large holes and unobserved regions remain unresolved. Early completion methods target single objects in normalized coordinates, while recent scene completion models typically complete within the given field of view and output point clouds or voxel grids rather than a complete mesh. The paper aims to recover a complete scene mesh from a partial point cloud while preserving observed geometry and scaling to scenes of different physical sizes.

AnchorComplete builds on TRELLIS.2 and takes a partial, colored point cloud as input. For partial-point conditioning, a point encoder with a frozen Utonia backbone and a trained four-layer Q-Former produces 768 anchor tokens that condition both flow stages through cross-attention, replacing image or text tokens. A local path voxelizes visible points, encodes them with the structure encoder, initializes visible latent cells with this evidence, and concatenates a visibility mask at every denoising step; Stage 2 also receives visible latents with a per-voxel visibility flag. Metric-scale coarse-to-fine generation uses level 0 as a single whole-scene cube and levels 1 and 2 as fixed-physical-size windows, with level-2 windows half as wide as level-1 windows and a 512^3 mesh grid of fixed cell size. Each finer window starts from its parent occupancy with alpha=0.8 in unobserved regions, windows overlap by a quarter of their width and keep only their centre, and the finest level merges decoded windows into a common grid for a single mesh. The two flow models are fine-tuned from TRELLIS.2 image-to-3D checkpoints with a rectified-flow velocity loss, training on window crops and parent predictions or damaged ground-truth parents.

• On complete-scene reconstruction at four views, AnchorComplete leads complete-region metrics on Replica and SCRREAM, with CD_all 0.067 and 0.076 and F0.1 0.819 and 0.786. §4.2
• Under sparse views, AnchorComplete leads CD_all and F-scores at every tested view count against DA3, NKSR, and Surflo; at 10 views its F0.05 is 0.829 vs 0.754 for NKSR on Replica and 0.697 vs 0.674 on SCRREAM. §4.3
• Ablating partial-point conditioning shows full conditioning gives CD_all 0.067, versus 0.105 without local injection and 0.086 without global tokens; removing local injection causes the largest drop. §4.4
• Coarse-to-fine levels improve high-curvature Fdet0.02 to 0.673 at level 2 from 0.429 at level 0, and F0.02 to 0.801 from 0.485. §4.4
• Without retraining, AnchorComplete remains effective across input point-cloud sources, with F0.05 ranging from 0.547 for VGGT-Omega to 0.761 for LiDAR. §A.4

• The paper states the current model predicts geometry only and does not model appearance; extending with RGB supervision and an appearance-aware decoder is left open.
• The paper states performance depends on upstream geometry quality, so drift, scale errors, or systematic depth bias may propagate to the generated mesh.
• The paper states training and evaluation focus on indoor scenes, leaving generalization to outdoor, unbounded, or substantially different scene distributions open.
• The paper states in the failure-case appendix that the model is trained only on clean geometry and treats the input as a reliable anchor, so severe upstream failures are inherited rather than corrected.

• Problem setting: input is a partial point cloud from reconstruction, SLAM, or scanning; output is a complete scene mesh, without requiring densely sampled posed images, oriented normals, or per-scene optimization. §1
• Global anchor tokens are encoded from point observations and fed to both TRELLIS.2 flow stages through cross-attention in place of image or text conditioning tokens. §3.2
• Local spatial injection voxelizes visible points, initializes visible latent cells with encoded visible geometry, and provides a visibility mask at every denoising step; Stage 2 also receives visible latents and a per-voxel visibility flag. §3.2
• Metric-scale coarse-to-fine generation decouples mesh cell size from scene extent: level 0 sees the whole scene, levels 1 and 2 use fixed-physical-size windows, and further levels would halve the window again. §3.3
• Parent warm start uses alpha=0.8 in unobserved regions, follows input geometry in observed regions, overlaps windows by a quarter of their width, keeps only centres, and merges decoded windows into one mesh at the finest level. §3.3
• Training fine-tunes two flow models from TRELLIS.2 image-to-3D checkpoints with a rectified-flow velocity loss; the parent for warm start is usually a prediction and otherwise a damaged ground-truth parent. §3.4
• Evaluation reports CD_vis, CD_all, and F-scores, with accuracy computed only inside input-view frustums because completion outside observed regions is multi-modal. §4.1
• On two real-world rooms from eight views shown qualitatively, NKSR, SG-NN, and Seen2Scene recover the layout but leave surfaces ragged and holed where coverage thins, while AnchorComplete closes them. §4.2