SUFLECA: Scaling Up Feature Learning for CAD-to-image Alignment

Recovering the 3D pose and shape of objects from a single image is a core challenge in robotics and augmented reality. Existing zero-shot methods often struggle with domain shifts, occlusions, and the high computational cost of iterative pose refinement. SUFLECA (Scaling Up Feature LEarning for CAD Alignment) addresses these issues by shifting from appearance-driven features to geometry-aware representations.

The authors introduce a two-part framework:

1. Large-scale Feature Learning: The model is trained on 674K images across 12 diverse datasets using Normalized Object Coordinates (NOCs) as a supervisory signal. By training on both real-world captures and synthetic CAD renders, the model learns compact, domain-invariant features.
2. Geometrically Consistent Matching: Instead of relying on simple nearest-neighbor matching, which is prone to noise, SUFLECA uses a robust correspondence estimation algorithm. It enforces mutual k-nearest-neighbor constraints and uses pairwise distance preservation to filter out geometrically inconsistent matches before performing final pose registration.

SUFLECA demonstrates significant improvements on the ScanNet25k benchmark. It achieves 33.4% category accuracy and 42.3% instance accuracy, outperforming the strongest zero-shot baseline by over 10 percentage points. Notably, SUFLECA is the first zero-shot method to surpass fully supervised approaches on this benchmark while maintaining a smaller computational footprint and sub-second inference time. The inclusion of a principled alignment quality score (S_fit) further allows for reliable ranking of estimated poses without requiring iterative refinement.

This work bridges the gap between zero-shot generalization and high-precision pose estimation. By moving away from costly, iterative render-and-compare optimization, SUFLECA provides a scalable and efficient solution for robotic perception tasks that require real-time, accurate object alignment in cluttered, real-world environments.