Hybrid Cross-Device Localization via Neural Metric Learning and Feature Fusion

2026-01-30

AI Summary

This work addresses the degradation in metric localization accuracy under cross-device scenarios caused by device heterogeneity and modality mismatch between map and query images. To tackle this challenge, the authors propose a dual-branch localization architecture that synergistically combines geometric and neural approaches: one branch integrates feature fusion with PnP solving, while the other leverages the MapAnything network for geometry-conditioned feedforward localization. The method introduces two key innovations—a neural-guided candidate map-frame pruning mechanism and a depth-conditioned optimization module—significantly enhancing both robustness and efficiency. Extensive experiments demonstrate state-of-the-art performance on the HYDRO and SUCCU benchmarks, with the approach achieving an outstanding score of 92.62% (R@0.5m, 5°) in the challenge evaluation.

Technology Category

• Intelligent Robots: Localization, Mapping, and Navigation
• Computer Vision: Multi-modal Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Systems and Infrastructure for Web, Mobile and WoT: Location- and context-aware Web and WoT applications and services
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

We present a hybrid cross-device localization pipeline developed for the CroCoDL 2025 Challenge. Our approach integrates a shared retrieval encoder and two complementary localization branches: a classical geometric branch using feature fusion and PnP, and a neural feed-forward branch (MapAnything) for metric localization conditioned on geometric inputs. A neural-guided candidate pruning strategy further filters unreliable map frames based on translation consistency, while depth-conditioned localization refines metric scale and translation precision on Spot scenes. These components jointly lead to significant improvements in recall and accuracy across both HYDRO and SUCCU benchmarks. Our method achieved a final score of 92.62 (R@0.5m, 5{\deg}) during the challenge.

Problem

• cross-device localization
• metric localization
• hybrid localization
• feature fusion
• neural metric learning

Innovation

• neural metric learning
• feature fusion
• cross-device localization
• MapAnything
• depth-conditioned localization