Answers are grounded in this paper and cite their sources.
Learning-based point cloud registration commonly trains local descriptors by pulling matched points together and pushing negatives apart. This is effective when negatives correspond to genuinely different local structures. However, many points in 3D scenes lie in low-distinctiveness regions, including table surfaces, walls, floors, object faces, and repeated structures. Treating such locally similar points as negatives forces the network to separate descriptors that should remain similar, thereby creating non-repeatable pseudo-salient features that do not generalize across scans. Existing methods improve correspondence quality through transformers, overlap prediction, coarse-to-fine matching, or outlier pruning, but descriptor supervision still largely relies on generic metric learning and does not determine where strong discrimination is geometrically reliable. This letter proposes a local-ambiguity-aware point cloud registration framework centered on graph-signal-processing (GSP) guided supervision allocation. The framework uses a KPConv backbone, computes a geometry reliability gate from local surface variation and normal-boundary response, models dense descriptors as graph signals, and obtains a low-pass reconstruction through superpoint coarsening based on farthest point sampling (FPS). The GSP residual and geometry gate redistribute descriptor variation: low-response regions receive weaker discriminative amplification, whereas middle- and high-response regions carry stronger circle-inspired metric supervision. Training concentrates discriminative learning on regions where high-quality, repeatable descriptors are more likely to form rather than manufacturing isolated feature peaks on ambiguous surfaces. A weighted variance regularizer prevents feature collapse. Experiments on 3DMatch and 3DLoMatch demonstrate improved indoor registration recall, while KITTI evaluation shows competitive outdoor performance with a lightweight KPConv-based architecture.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: