Title: Z-FLoc: Zero-Shot Floorplan Localization via Geometric Primitives

Abstract:

Estimating a camera’s position relative to a pre-established map is a core challenge in computer vision, known as visual localization. Floorplans offer a compelling mapping solution due to their widespread availability, compact nature, and resistance to changes in visual appearance. Nevertheless, aligning camera imagery with the geometric structure of floorplans presents a significant hurdle. Current solutions rely on data-driven learning, which necessitates extensive training datasets and retraining for specific settings, thereby hindering real-world applicability. To overcome these limitations, we introduce Z-FLoc, a zero-shot localization technique capable of generalizing to new environments without requiring any model retraining. Our approach is grounded in the observation that lines and circles, as dominant geometric primitives, are prevalent in constructed environments and serve as robust, appearance-independent structural cues. We derive these primitives from bird’s-eye-view (BEV) projections of monocular 3D reconstructions and align them with floorplans using specialized minimal solvers embedded in a robust estimation framework. Evaluations on both synthetic and real-world datasets demonstrate that our method surpasses existing learning-based state-of-the-art techniques in unseen scenarios, consistently utilizing a single, static set of hyperparameters. The project’s source code will be released publicly.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC