Line up a CAD model with an object in a single photo and you can hand a robot or an AR system its position, orientation, and size, nine numbers in all. Recent zero-shot methods match image patches to the model using visual foundation features, but those matches lean on appearance and fall apart under occlusion or the sim-to-real gap.

SUFLECA leans on geometry instead. It trains compact, geometry-aware features using Normalized Object Coordinates supervision across 674,000 images from a dozen datasets, then pairs them with a matcher that enforces clean one-to-one correspondences. Each object aligns in under a second, no iterative refinement. On ScanNet25k it reports 33.4 and 42.3 percent category and instance accuracy, ahead of the best zero-shot baseline and, the authors say, past fully supervised methods for the first time on that benchmark.

Accuracies in the 30s and 40s show this remains hard, so the paper and code fill in the rest.