Skip to content
Book Open access

ReasonLoc: Geometric-Language Localization in 3D Point Clouds via Spatial Reasoning

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 41 references

Abstract

Text-based localization in 3D point clouds aims to ground free-form natural-language descriptions by reasoning over object geometry and spatial relations. Existing text-to-localization methods have shown promising performance, but they remain limited by ambiguous template-based supervision and insufficient modeling of explicit geometric relations in 3D visual representations. To address these challenges, we propose ReasonLoc, a unified spatial reasoning framework that strengthens supervision, representation, and geometric language understanding for text-guided 3D localization. Specifically, we first quantify and reduce spatial ambiguity by extending directional supervision from 5 to 9 categories, leading to an updated benchmark (KITT360-9). Second, we introduce a Spatial Information Module (SIM) that constructs 5D geometric relation graphs and injects them into a query-conditioned Relation Reranker for coarse retrieval, followed by a dual-channel geometric fusion network for fine-grained coordinate regression. Third, we employ a large language model (LLM) as a constrained spatial reasoner to move beyond rigid templates, which converts free-form descriptions into structured directional and geometric cues. Finally, we establish a new free-form evaluation setting via top-candidate differencing to assess robustness under natural language variation. Extensive experiments demonstrate that ReasonLoc improves geometric-language grounding in 3D point clouds, achieving up to 31% gains in Top-1 retrieval and 33% gains in Top-1 localization over existing methods, while maintaining competitive zero-shot transfer on Toronto3D and Paris-Lille.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.