CAOA boosts CAD alignment accuracy by 17% on real-world scans

Researchers from the University of Texas at Dallas present CAOA (Completion-Assisted Object-CAD Alignment), a novel method that significantly improves the accuracy of aligning 3D CAD models to objects in real-world RGB-D scans. The core challenge is estimating a 9-Degree-of-Freedom (DoF) pose—position, rotation, and scale—from noisy, incomplete scans that often suffer from segmentation errors. CAOA tackles this with two key innovations: a semantically and contextually aware point cloud completion module that fills in missing geometry, and a symmetry-aware relative pose estimation algorithm that uses a dedicated loss function to handle symmetric ambiguities.

On the Scan2CAD benchmark, CAOA achieves a 17% accuracy improvement over state-of-the-art methods. To bridge the synthetic-to-real domain gap that plagues existing completion models, the team introduces a tailored synthetic data generation strategy for indoor scenes and releases S2C-Completion, a new expert-annotated dataset of over 8,500 object-CAD pairs from Scan2CAD. This dataset is designed as a benchmark for real-world indoor single-object completion, validated through quantitative comparisons with widely used completion datasets. The code is available on GitHub.