SUFLECA zero-shot CAD alignment tops supervised on ScanNet25k

A quiet result out of Luxembourg is worth a look if you build anything that has to place a CAD model in the same coordinate frame as a real photo. The SUFLECA paper, from a team at the University of Luxembourg's SnT centre and Universidad de Zaragoza, reports the first zero-shot method to surpass fully supervised baselines on ScanNet25k, a standard benchmark for aligning known 3D shapes to single RGB images.

The specific claim is that SUFLECA reaches 33.4% category-averaged and 42.3% instance-averaged accuracy on the ScanNet25k NMS protocol. That is 10.3 and 12.2 percentage points above ZeroCAD, the previously strongest zero-shot baseline, and it also outperforms MultiObj-SPARC, which was trained with direct 9D pose supervision on ScanNet. The authors get there without any pose-annotated data by scaling Normalized Object Coordinates supervision to 674K images drawn from 12 real and synthetic datasets, versus the roughly 300K single-object synthetic images across nine categories used in prior zero-shot work.

Why the shift matters for people building robotics stacks or AR pipelines: CAD alignment is usually a bottleneck because collecting real 9D pose labels is expensive, and previous zero-shot attempts leaned on iterative render-and-compare loops that were slow and memory-hungry. SUFLECA claims sub-second per-object alignment without iterative refinement, and the authors report the smallest GPU memory footprint among zero-shot methods in their headline comparison. Code is released at github.com/snt-arg/SUFLECA under CC BY 4.0.

The honest caveat sits inside the paper itself. Alignment quality is still capped by the CAD retrieval step; when the authors swap ground-truth Scan2CAD annotations for a fully zero-shot retrieval pipeline built on GroundedSAM plus OSCAR, instance retrieval accuracy collapses to 3.8% and alignment accuracy drops more than 20 percentage points. The method is also currently restricted to indoor scenes and common object categories, and the reporting does not give you numbers on training compute or a breakdown of how the pipeline behaves under heavy occlusion.

For teams already shipping supervised alignment via ROCA-style pipelines, the practical question is whether SUFLECA-scale weakly-supervised training is now the cheaper path to a working stack, especially as open-set CAD retrieval improves.