VimGeo: Efficient Cross-View Geo-Localization with Vision Mamba Architecture

Jinglin Huang, Maoqiang Wu, Peichun Li, Wen Wu, Rong YuIJCAI 2025 · DOI

TL;DR

This paper introduces VimGeo, an efficient cross-view geo-localization system that leverages the Vision Mamba (Vim) architecture. By replacing the traditional classification head with a novel Channel Group Pooling (CGP) module and introducing a Dynamic Weighted Batch-tuple Loss (DWBL), VimGeo achieves state-of-the-art accuracy while significantly reducing computational complexity and model parameters.

What Problem Does This Paper Address?

Cross-view geo-localization, the task of determining the geographic location of a ground-level image by matching it with a geo-tagged aerial image, is crucial for applications like autonomous driving and robot navigation. However, this task is inherently challenging due to significant variations in viewpoints, appearance, resolution, and perspective between ground and aerial images. Existing methods, while improving, often suffer from high model complexity, excessive resource consumption, and difficulties in optimizing with challenging samples. This paper aims to address these limitations by proposing a more efficient and accurate architecture.

Methods

The core of VimGeo is the Vision Mamba (Vim) architecture, chosen for its efficient state-space model (SSM) approach that avoids the high computational cost of self-attention mechanisms found in Transformers. The authors introduce two key innovations:

  • Channel Group Pooling (CGP): This module replaces the traditional classification head. Instead of a direct classification, CGP partitions feature maps into channel groups, applies average pooling within each group, and then merges these features. This allows for efficient global feature integration and flexible control over output dimensions without increasing model parameters or computational overhead.
  • Dynamic Weighted Batch-tuple Loss (DWBL): This novel loss function is designed to improve optimization by dynamically adjusting the weighting of negative samples. It focuses more on "hard" negative samples (those visually similar to the anchor but incorrect) by assigning them higher weights. This mechanism helps refine discriminative power and improve the model's ability to achieve accurate geo-localization, especially in challenging scenarios with subtle visual differences.

Datasets

The proposed VimGeo model is evaluated on three widely-used cross-view geo-localization datasets:

  • CVUSA: Consists of suburban landscapes with 35,532 training and 8,884 testing pairs.
  • CVACT: Focuses on urban regions in Canberra, with 35,532 training, 8,884 validation, and 92,802 testing pairs.
  • VIGOR: A larger dataset with 105,214 ground images and 90,618 aerial images, allowing for evaluation on both "Same-area" and "Cross-area" protocols.

Key Insights and Ideas

  • Vision Mamba's Efficiency: The adoption of the Vision Mamba architecture proves to be a strong foundation, offering a good trade-off between computational efficiency and performance, especially compared to Transformer-based methods.
  • CGP for Feature Integration: Replacing the classification head with CGP is a clever way to achieve efficient global feature fusion without the parameter bloat of fully connected layers. This highlights the importance of how features are aggregated after extraction.
  • DWBL for Hard Negatives: The dynamic weighting of negative samples in DWBL is crucial for improving the model's ability to distinguish between very similar but incorrect candidates. This is a key insight into overcoming the "hard negative" problem in retrieval tasks.
  • Complementary Components: The ablation studies clearly show that both CGP and DWBL contribute significantly and complement each other, leading to the best overall performance when used together.

Key Results

VimGeo achieves state-of-the-art performance on several benchmarks:

  • On CVUSA, it achieves an R@1 of 96.19%.
  • On CVACT_val, it achieves an R@1 of 87.62%.
  • On CVACT_test, it significantly outperforms other methods with an R@1 of 81.69%, which is 9.54% higher than the next best method (FRGeo).
  • VimGeo also demonstrates superior computational efficiency, achieving the best computational complexity (11.023 GFLOPs) and requiring only 50.868M trainable parameters while maintaining high accuracy.

Limitations and Future Work

  • VIGOR Dataset Performance: The paper notes that the DWBL loss was found to be unsuitable for the VIGOR dataset, requiring a switch to the Weighted (B+1)-tuple Loss (WBL). This suggests that the optimal loss function might be dataset-dependent or that further refinement of DWBL for complex datasets like VIGOR is needed.
  • Future Work: While not explicitly detailed, the success of VimGeo suggests future work could explore:
  • Further optimization of the CGP module for different feature extraction backbones.
  • Investigating variations of DWBL to adapt to diverse dataset characteristics.
  • Applying VimGeo to other related tasks that involve matching images from different viewpoints.

Summary updated

BibTeX
@inproceedings{huang2025vimgeo,
  author = {Jinglin Huang and Maoqiang Wu and Peichun Li and Wen Wu and Rong Yu},
  title = {{VimGeo: Efficient Cross-View Geo-Localization with Vision Mamba Architecture}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/133},
}
Figure