Answers are grounded in this paper and cite their sources.
Computational evaluation demonstrates improved coarse registration recall in low-overlap 3D point clouds, indicating that dummy-node partial assignment reduces false correspondences.
Reliable coarse registration remains challenging when point-cloud scans have limited overlap and some keypoints lack valid correspondences. Conventional descriptor matching selects correspondences independently, which can produce many-to-one matches and force keypoints from non-overlapping regions into incorrect correspondences. To address this problem, we propose a Hypothesis-Feedback Partial Assignment framework (HFPA) for coarse registration. HFPA constructs candidate correspondences from ISS keypoints and FPFH descriptors and provides each source keypoint with an unmatched option through a source-specific dummy node. An augmented Hungarian assignment jointly determines real matches and rejection states while enforcing one-to-one constraints on real target keypoints. Correspondence residuals under rigid-transformation hypotheses are then fed back to update the matching and rejection costs. The correspondences are iteratively refined, and the final coarse transformation is estimated using RANSAC. On indoor scan pairs with different overlap levels, HFPA achieved the same registration recall as Full Hungarian but a lower mean rRMSE. On the eight-scene low-overlap subset, HFPA achieved a registration recall of 62.5%, compared with 37.5% for Full Hungarian and 50.0% for FPFH + NNDR. These results show that source-specific dummy nodes and hypothesis-residual feedback reduce forced incorrect correspondences and improve the reliability of conventional local-descriptor pipelines for low-overlap coarse registration.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: