GeoFlow: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction

A lightweight, real-time framework for fine-grained cross-view geolocalization. Accepted to CVPR 2026.

This project introduces a lightweight and real-time framework for fine-grained cross-view geolocalization, designed to overcome the accuracy–efficiency trade-off that limits practical deployment.

Problem

Fine-grained cross-view geolocalization (FG-CVG) requires precise localization while operating under strict runtime constraints.
Existing approaches often trade accuracy for speed or rely on heavy iterative inference, making them unsuitable for real-time applications such as robotics, autonomous navigation, and augmented reality.

Key Idea

We decouple expensive visual reasoning from fast geometric refinement:

  1. Single-pass visual encoding:
    Ground and satellite images are processed once using lightweight CNN backbones and a cross-attention module to extract a shared visual representation.

  2. Fast iterative refinement:
    Localization is performed by iteratively refining multiple pose hypotheses using a compact MLP, enabling robustness without repeatedly running the visual backbone.

This design enables iterative and uncertainty-aware localization while remaining efficient for real-time use.

Technical Details

This framework was implemented with a strong emphasis on efficiency, modularity, and real-time deployment:

  • Framework: PyTorch
  • Visual encoders: Lightweight CNN backbones (e.g., EfficientNet, ConvNet-style architectures)
  • Cross-view fusion: Cross-attention mechanism for aligning ground-level and satellite feature representations
  • Localization head: Small MLP predicting probabilistic displacement (distance and direction)
  • Inference strategy: Multi-hypothesis iterative refinement with shared visual features
  • Training: End-to-end supervised learning with regression-based objectives
  • Benchmarks: Evaluated on standard FG-CVG datasets including KITTI and VIGOR

The architecture is designed such that computationally expensive components are executed only once, while refinement operates entirely in a low-dimensional latent space.

Results

The proposed framework achieves compelling localization accuracy on benchmarks such as KITTI and VIGOR, while maintaining strong runtime efficiency suitable for real-time applications (~30 FPS(Frame Per Second)).

References