SUFLECA: Scaling Up Feature Learning for CAD-to-image Alignment
SUFLECA reports zero-shot CAD alignment that beats prior zero-shot baselines on ScanNet25k while running sub-second per object.
The paper targets 9D pose estimation from a single RGB image: rotation, translation, and anisotropic scale. It trains geometry-aware features with Normalized Object Coordinates across up to 12 real and synthetic datasets, then uses a matching algorithm meant to keep CAD-to-image correspondences geometrically consistent. On ScanNet25k, it reports 32.8% category accuracy and 42.6% instance accuracy, ahead of the strongest zero-shot baseline by 9.7 and 12.5 percentage points. Code is listed as available in the paper. HF Daily Papers' note
The paper targets 9D pose estimation from a single RGB image: rotation, translation, and anisotropic scale. It trains geometry-aware features with Normalized Object Coordinates across up to 12 real and synthetic datasets, then uses a matching algorithm meant to keep CAD-to-image correspondences geometrically consistent. On ScanNet25k, it reports 32.8% category accuracy and 42.6% instance accuracy, ahead of the strongest zero-shot baseline by 9.7 and 12.5 percentage points. Code is listed as available in the paper. HF Daily Papers' note
score 4