Space-filling curves (SFCs) are a lightweight way to accelerate nearest-neighbor search in 3D point clouds by mapping points into a one-dimensional order. Yet the order induced by the curve can disagree with true Euclidean distances, creating a gap that may mislead models trained on local neighborhoods.

To address this, Konan Sasaki, Junya Komori, and Kazuhiro Hotta propose a loss function that measures distances between adjacent points along the SFC and pushes the model to align features for pairs that are close on the curve but far apart in 3D space. This directly counteracts the curve's spurious neighbor relationships.

The authors integrated the loss into an existing high-performing segmentation model and evaluated it on three datasets. Their method beat conventional baselines on both indoor and outdoor point cloud benchmarks, suggesting that explicitly closing the SFC-spatial gap is broadly useful. The work appears in an ECCV 2026 workshop paper on scalable scene generation and geometric scene understanding.