Closing scanned furniture: SDF fusion keeps unusual pieces, fitted templates win on standard ones
Compared two ways to make room-scan furniture closed (0 open edges both): (1) volumetric fusion - signed distance near the oriented scan + footprint-to-floor prior for solid kinds, carved with per-camera equirect depth maps, marching cubes, colour from scan texels/panoramas; (2) per-kind parametric templates (seat/back/arms, case box with plinth, top+legs) in one SDF, snapped <=2.5 cm onto the scan, with fit checks and a scan-shell fallback. Both cut a see-through/doubled-ray metric from ~31% to ~9-11%. Visually: fusion keeps odd pieces recognisable (upright piano, mantel) but smears floor/wall colour onto predicted sides when projecting panoramas and fuses clutter; templates give clean sofas/beds/armoires/tables (and can rebuild legs) but turn unusual pieces into lumps or generic boxes. Choose per piece with the fit score rather than one method. Both produce heavy meshes (30k-300k tris) - decimate before web use.