RegionUDF: Region-Aware Unsigned Distance Fields for Surface Reconstruction from Point Clouds

Xin Deng, Bo Yang, Yifei Shi, Bing WANG

Keywords: Surface Reconstruction, Unsigned Distance Filelds, Point Clouds

TL;DR: We propose a discriminative query-region representation that improves UDF learning and enables high-quality open-surface reconstruction.

Abstract: Distance fields offer a powerful representation for continuous geometry, yet current learning-based neural unsigned distance fields (UDFs) remain limited in their ability to capture data patterns and generalize to real-world open surfaces. Point-Based methods mitigate grid quantization errors but current work often oversmooth local details, as query features are obtained solely through interpolation of point-wise features which are aggregated over large receptive fields. To address this, we propose a $ \textit{discriminative region representation} $ that fuses narrow neighborhood features with broader contextual point-wise features, and a $ \textit{primitive-based region representation} $ that decomposes the query region into triplet-defined primitives, enabling the detailed encoding of local surface geometry and the clear distinction of multi‑layer structures. Building on these designs, we propose $ \textit{RegionUDF} $, a region-aware UDF framework that achieves state-of-the-art open-surface reconstruction on both object- and room-level scenes, with additional validation on watertight shapes. Extensive experiments on synthetic and real-world datasets demonstrate superior accuracy and robust cross-domain generalization. Our source code will be available at $ \textit{[no-name-for-blind-review]} $.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 11963