Closing torn scan meshes: MLS signed distance band + solid prior + camera space carving, then marching cubes
For furniture cut from a photogrammetry/Matterport mesh (holes, inconsistent winding, single-sided shells): (1) orient sample normals by which scan cameras see each triangle (per-camera panoramic depth buffer of the whole scene: splat points far-to-near, last write wins, 3x3 min filter), unseen triangles copy nearby oriented neighbours; (2) voxelize (~1.5 cm), inside a band tau=3.5 voxels use MLS f(x)=sum w (x-p).n / sum w with gaussian w (fills small holes; behind a one-sided surface it yields a closed slab of thickness tau); (3) beyond the band use a shape prior only for solid pieces (footprint from floor to scanned top), and carve prior voxels visible from >=2 cameras (closer than the depth buffer minus 5 cm); (4) OR in the eroded prior core so scan junk cannot open pockets, binary_closing(2) to seal thin cracks, binary_fill_holes, drop tiny components, gaussian(1 voxel), skimage marching_cubes; check winding with the field gradient. Result: 0 open edges, see-through/doubled from ~31% to ~11%. Note: ray-layer metrics that bin depth (~6 mm) count thin parts (<~2 cm) in front of other surfaces and narrow gaps between cushions as doubled layers even on a closed mesh; closing narrow gaps helps a lot. Colour: keep scan texture on vertices within 1 voxel of the scan (sharper than reprojecting 2k equirect panos), use panorama projection only for predicted surfaces, then Laplacian-smooth colours off the scan to remove speckle.