[Paper Review] Visual and Object Geo-localization: A Comprehensive Survey

Daniel J. Wilson, Xiaohan Zhang+2 more|arXiv (Cornell University)|Dec 30, 2021

Advanced Image and Video Retrieval Techniques8 citations

TL;DR

This comprehensive survey provides a detailed analysis of visual geo-localization, covering image and object geo-localization using deep learning and cross-view approaches. It evaluates state-of-the-art methods, benchmarks on public datasets, and highlights the growing importance of cross-view and transformer-based models in improving localization accuracy and robustness.

ABSTRACT

The concept of geo-localization refers to the process of determining where on earth some `entity' is located, typically using Global Positioning System (GPS) coordinates. The entity of interest may be an image, sequence of images, a video, satellite image, or even objects visible within the image. As massive datasets of GPS tagged media have rapidly become available due to smartphones and the internet, and deep learning has risen to enhance the performance capabilities of machine learning models, the fields of visual and object geo-localization have emerged due to its significant impact on a wide range of applications such as augmented reality, robotics, self-driving vehicles, road maintenance, and 3D reconstruction. This paper provides a comprehensive survey of geo-localization involving images, which involves either determining from where an image has been captured (Image geo-localization) or geo-locating objects within an image (Object geo-localization). We will provide an in-depth study, including a summary of popular algorithms, a description of proposed datasets, and an analysis of performance results to illustrate the current state of each field.

Motivation & Objective

• To provide a comprehensive, up-to-date survey of visual geo-localization, including image and object-level localization.
• To analyze the evolution and current state of deep learning-based approaches in geo-localization, especially cross-view and transformer-based methods.
• To systematically compare existing algorithms, datasets, and evaluation metrics across single-view, cross-view, and object geo-localization subfields.
• To identify key challenges and limitations in current methods, such as reliance on handcrafted features and GAN instability.
• To guide future research by highlighting underexplored areas, including self-supervised learning and robustness to real-world data variations.

Proposed method

• Categorizes visual geo-localization into three main subfields: single-view, cross-view, and object geo-localization.
• Reviews key techniques including Siamese networks for feature learning, graph-based models for landmark connectivity, and GANs for synthetic image generation.
• Analyzes tracker-based, re-identification, and triangulation methods for object geo-localization, emphasizing their strengths and limitations.
• Evaluates performance using standard metrics such as top-k accuracy, mean average precision (mAP), and recall at specific distances.
• Compares methods across public datasets like UBER-NET, Mapillary, and GPS-RetinaNet, highlighting differences in data collection and evaluation protocols.
• Highlights the role of modern architectures like vision transformers and self-supervised learning in improving feature representation and generalization.

Experimental results

Research questions

• RQ1 What are the core differences and trade-offs between single-view, cross-view, and object geo-localization approaches?
• RQ2 How do deep learning-based methods, particularly Siamese networks and GANs, improve upon traditional handcrafted feature methods in geo-localization?
• RQ3 What are the performance characteristics and limitations of tracker-based, re-identification, and triangulation methods in object geo-localization?
• RQ4 How do evaluation metrics and dataset design influence reported performance in geo-localization research?
• RQ5 What future directions—such as self-supervised learning and vision transformers—hold the most promise for advancing the field?

Key findings

• Cross-view geo-localization methods, which match ground-view queries to satellite reference images, achieve state-of-the-art performance, with mAP values exceeding 0.92 on the UBER-NET dataset.
• Tracker-based object geo-localization methods like GPS-RetinaNet achieve a top-10 accuracy of 5.81 meters on the UBER-NET dataset, demonstrating high precision in object-level localization.
• Re-identification-based approaches such as GeoGraph and Mapillary TLG achieve mAP scores of 0.924 and 0.882 respectively, showing strong performance when multiple images of the same object are available.
• Triangulation-based methods like MRF achieve high precision (95th percentile error of 1.89m for traffic lights), but require accurate camera parameters and are sensitive to depth estimation errors.
• Generative models using GANs show promise in synthesizing realistic ground-view images for training, though they suffer from training instability and mode collapse.
• Vision transformers and self-supervised learning are emerging as key enablers for improved feature learning, with potential to outperform convolutional networks in future geo-localization systems.