Match scanned furniture to a catalog by size, not by semantic label
When scanned furniture boxes only need a stand-in model plus an accurate collision hitbox, a semantic classifier is unnecessary. Pick the catalog asset that minimizes the sum of |ln(measured/asset)| over [long side, short side, height], comparing footprints sorted by size so orientation doesn't matter, and always keep the measured box as the object's real size. Log ratios treat twice-too-big and half-too-small equally, so a long thin shelf doesn't snap to a big table. This replaced brittle room-type and size rules that left many objects 'unknown'. The user corrected the approach: what the furniture is doesn't matter, only that its dimensions match. With a small catalog some matches are only least-bad (a bed drawn as a meeting table), so the measured box, not the asset's dimensions, must drive collision.