SUFLECA: Better CAD-to-Image Alignment

Summary & Methodology Analysis

SUFLECA addresses the challenge of zero-shot CAD-to-image alignment, where existing appearance-driven foundation models suffer from domain gaps and produce geometrically inconsistent correspondences that require costly iterative refinement. The paper aims to provide more robust and accurate alignment without requiring specific training for each new object. The paper's novel approach scales up NOC-supervised feature learning, which trains models to learn feature representations guided by Normalized Object Coordinate maps, on 674K images across 12 datasets. This produces geometry-aware feature representations. Additionally, it implements a novel correspondence estimation algorithm that replaces semantic-only matching with mutual k-nearest-neighbors matching, filtered by a geometric consensus. For standard processing steps, SUFLECA uses an existing monocular metric depth estimator to lift image pixels to 3D points, and RANSAC-based Procrustes registration is used for final 9D alignment. An alignment quality score (S_fit) is computed based on the log-determinant of the information matrix of registration residuals. On the ScanNet25k benchmark, SUFLECA achieves 33.4% category accuracy and 42.3% instance accuracy, outperforming the strongest zero-shot baseline by 10.3 and 12.2 percentage points respectively. However, a key limitation is that alignment quality depends on the accuracy of CAD model retrieval from occluded real-world observations. The method is also currently restricted to indoor scenes and common object categories.