RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs

2026-07-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Retrieval-In-Matching (RIM) is a novel framework for global visual localization of Unmanned Aerial Vehicles (UAVs) in GNSS-challenged environments. It tackles significant cross-domain appearance and viewpoint shifts between UAV and remote-sensing reference imagery by sampling UAV-viewpoint reference views from Google 3D Tiles. The core innovation is a shared-representation architecture where a frozen DINOv2-B retriever, adapted via a two-stage fine-tuning of SALAD, also distills a local-descriptor decoder. This allows a single DINOv2-B forward pass to handle both global retrieval and local description, eliminating a second foundation-model backbone. Evaluated zero-shot on EPFL Urbanscape and Chang'an Park datasets, RIM improved Recall@1 over SALAD by 8.55/13.77 percentage points at 25/50m on EPFL and 4.45/8.94 points on Park. A complete localization query (Top-K=5) runs in 67.9 ms, making it 1.8x faster than the strongest sparse-matching baseline and over 40x faster than RoMa, while maintaining re-ranking accuracy.

Key takeaway

For robotics engineers developing UAV navigation systems in GNSS-challenged environments, RIM presents a compelling solution. You should consider integrating this shared-representation architecture to achieve robust 6-DoF visual localization. Its 1.8x speed improvement over sparse-matching baselines, while maintaining re-ranking accuracy, allows for efficient real-time operation without sacrificing precision. This enables more reliable cold-start re-localization for SLAM tracking recovery.

Key insights

Retrieval-In-Matching (RIM) efficiently localizes UAVs by sharing a single foundation model backbone for both global retrieval and local matching.

Principles

• Pre-align viewpoints to reduce geometric discrepancy.
• Adapt foundation models with pose-near positives and hard negatives.
• Reuse foundation model tokens for both global retrieval and local matching.

Method

Construct UAV-viewpoint 3D Tiles database. Adapt SALAD via two-stage fine-tuning. RIM then shares DINOv2-B token field for global retrieval and distilled local-descriptor decoding.

Topics

• UAV Localization
• Visual Place Recognition
• DINOv2
• 3D Tiles
• Cross-Domain Adaptation
• Real-time Inference

Best for: Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs — A Retrieval-In-Matching framework efficiently localizes UAVs in GNSS-challenged environments by reusing a single DINOv2-B backbone for both global retrieval and local description.
• OffNadirLoc: Benchmark and Framework for Challenging UAV-to-Satellite Geo-Localization under Large Off-Nadir Views — Large off-nadir UAV-to-satellite geo-localization requires structure-aware feature weighting and multi-view coherent learning for robustness.
• G2IA: Geometry-Guided Instance-Aware Retrieval and Refinement for Cross-Modal Place Recognition — G2IA improves cross-modal place recognition by combining geometry-guided retrieval with instance-aware refinement, overcoming modality gaps and perceptual aliasing.
• Visual Relocalization from Sparse Views in Aliased and Low-Texture Environments via Novel View Synthesis — Combining MVS and LiDAR depths in 3DGS training significantly improves visual relocalization accuracy in extreme environments.
• GAP-GDRNet: Geometry-aware monocular 6D pose estimation for spacecraft using synthetic geometric supervision — GAP-GDRNet enhances monocular 6D spacecraft pose estimation by integrating attention-based feature refinement and patch-level self-attention into the GDR-Net pipeline.
• Compact Object-Level Representations with Open-Vocabulary Understanding for Indoor Visual Relocalization — OpenReLoc uses object-level representations and open-vocabulary semantics for robust indoor visual relocalization.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.