Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learnable Query Aggregation with KV Routing for Cross-view Geo-localisation

Hualin Ye, Bingxi Liu, Jixiang Du, Yu Qin, Ziyi Chen, Hong Zhang

TL;DR

This paper tackles cross-view geo-localisation (CVGL) by addressing severe viewpoint and scale shifts that hinder robust feature aggregation. It introduces a lightweight, parameter-efficient CVGL framework built on a DINOv2 backbone with convolutional adapters, a Multi-scale Channel Reallocation (MSCR) module for multi-scale spatial understanding, and a MoE-enhanced aggregation with KV routing to dynamically represent domain-specific cues. The approach achieves competitive results on University-1652 and SUES-200 datasets with significantly fewer trainable parameters, and ablations confirm the benefits of each component, especially the MoE KV routing. Overall, the method offers a practical, scalable solution for accurate CVGL in diverse real-world settings, enabling robust geo-localisation without heavy fine-tuning of foundation models.

Abstract

Cross-view geo-localisation (CVGL) aims to estimate the geographic location of a query image by matching it with images from a large-scale database. However, the significant view-point discrepancies present considerable challenges for effective feature aggregation and alignment. To address these challenges, we propose a novel CVGL system that incorporates three key improvements. Firstly, we leverage the DINOv2 backbone with a convolution adapter fine-tuning to enhance model adaptability to cross-view variations. Secondly, we propose a multi-scale channel reallocation module to strengthen the diversity and stability of spatial representations. Finally, we propose an improved aggregation module that integrates a Mixture-of-Experts (MoE) routing into the feature aggregation process. Specifically, the module dynamically selects expert subspaces for the keys and values in a cross-attention framework, enabling adaptive processing of heterogeneous input domains. Extensive experiments on the University-1652 and SUES-200 datasets demonstrate that our method achieves competitive performance with fewer trained parameters.

Learnable Query Aggregation with KV Routing for Cross-view Geo-localisation

TL;DR

This paper tackles cross-view geo-localisation (CVGL) by addressing severe viewpoint and scale shifts that hinder robust feature aggregation. It introduces a lightweight, parameter-efficient CVGL framework built on a DINOv2 backbone with convolutional adapters, a Multi-scale Channel Reallocation (MSCR) module for multi-scale spatial understanding, and a MoE-enhanced aggregation with KV routing to dynamically represent domain-specific cues. The approach achieves competitive results on University-1652 and SUES-200 datasets with significantly fewer trainable parameters, and ablations confirm the benefits of each component, especially the MoE KV routing. Overall, the method offers a practical, scalable solution for accurate CVGL in diverse real-world settings, enabling robust geo-localisation without heavy fine-tuning of foundation models.

Abstract

Cross-view geo-localisation (CVGL) aims to estimate the geographic location of a query image by matching it with images from a large-scale database. However, the significant view-point discrepancies present considerable challenges for effective feature aggregation and alignment. To address these challenges, we propose a novel CVGL system that incorporates three key improvements. Firstly, we leverage the DINOv2 backbone with a convolution adapter fine-tuning to enhance model adaptability to cross-view variations. Secondly, we propose a multi-scale channel reallocation module to strengthen the diversity and stability of spatial representations. Finally, we propose an improved aggregation module that integrates a Mixture-of-Experts (MoE) routing into the feature aggregation process. Specifically, the module dynamically selects expert subspaces for the keys and values in a cross-attention framework, enabling adaptive processing of heterogeneous input domains. Extensive experiments on the University-1652 and SUES-200 datasets demonstrate that our method achieves competitive performance with fewer trained parameters.
Paper Structure (25 sections, 16 equations, 4 figures, 6 tables)

This paper contains 25 sections, 16 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Cross-view Fine-Tuning based on Convolutional Adapters
  4. Multi-scale Channel Reallocation
  5. Multi-scale Feature Extraction
  6. Residual Signal Modulation
  7. Final Output
  8. MoE-enhanced Aggregation
  9. BoQ-like Aggregation
  10. Mixture-of-Experts for Keys and Values
  11. Loss Function
  12. Experiment
  13. Experimental Settings
  14. Dataset
  15. Implementation Details
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Each column presents images from the same location. The top row shows satellite imagery, while the second and third rows depict UAV-captured views. UAV images from a single location often exhibit substantial variations in scale and viewpoint, whereas images from different locations tend to contain highly diverse discriminative structures (e.g., roads and buildings). By leveraging specialised expert models, the MoE-enhanced aggregation layer selectively activates the most suitable experts, thereby improving performance on these challenging cases.
  • Figure 2: The proposed system is illustrated above, comprising three contributions: (1) convolution-based adaptation, (2) the MSCR module, and (3) an MoE-enhanced aggregator. The right-hand side provides architectural details of the first two components, while the structure of the aggregator is presented on the following page.
  • Figure 3: Detailed illustration of the BoQ module with KV routing.
  • Figure 4: Feature map visualisations for the ablation study from two scenes. From left to right: the original input image, w/o PEFT (fully frozen DINOv2), w/ PEFT (DINOv2 fine-tuned via Convolution), and our method (full model incorporating both Convolution fine-tuning and the MSCR module).