AnchorComplete: Recovering Complete Scene Meshes from Partial Point Clouds
Weirong Chen, Ismail El-Yamany, Ganlin Zhang, Chuanxia Zheng, Andrea Vedaldi, Daniel Cremers
Technical University of Munich · Munich Center for Machine Learning · Nanyang Technological University · University of Oxford
Published November 2, 2026 · on the feed November 5, 2026 · ▲ 11 · arXiv 2611.03228
The 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.
The Brief
1:47A short podcast summary of the paper. Written and voiced by AI.
Transcript
AnchorComplete. Recovering Complete Scene Meshes from Partial Point Clouds. By Weirong Chen and five others, from the Technical University of Munich, the Munich Center for Machine Learning, and two other institutions. Published on the second of November.
Most image-based 3D reconstructors recover what the views actually reached: the visible surfaces. What comes out is a partial point cloud with large holes in it, and the geometry nobody observed is simply left unresolved. Any surface that no view happened to see stays missing. For a whole scene, rather than a single object seen from every side, that leaves a reconstruction that is accurate where it exists and absent everywhere else.
AnchorComplete starts from that partial point cloud and reconstructs a complete scene mesh from it. Two requirements are held at once. The geometry that was observed has to be preserved, not redrawn, and the framework has to support scenes of varying scales. It builds on TRELLIS.2 and adds two pieces to it.
The first is partial-point conditioning. It integrates global scene tokens with local spatial evidence, and its purpose is to anchor generation to the input, so that what gets generated stays tied to the points that were really there. The second is metric-scale coarse-to-fine generation. Scene geometry is refined progressively inside fixed-size windows, while global structural consistency is maintained across the whole scene. The windows keep each refinement local; the global consistency keeps the result one scene rather than a patchwork.
Experiments run on Replica and SCRREAM. On both, AnchorComplete consistently outperforms strong baselines in complete-scene reconstruction, and the margin is described as most pronounced under sparse-view settings, where the fewest views feed the input and the most is left to complete.
The method is defined as much by what it leaves alone as by what it adds. Observed geometry stays as observed, and generation fills in the rest, held to the evidence instead of free to replace it.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free