Large-Scale Image Geo-Localization Using Dominant Sets.

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Geo-localization using image matching is a complex problem, especially in large-scale urban environments.
• Existing methods often rely on heuristic voting or fixed parameters, limiting their robustness and efficiency.

Purpose of the Study:

• To develop a more accurate and efficient geo-localization framework using image feature clustering.
• To overcome limitations of existing methods by employing a dynamic and robust clustering approach.

Main Methods:

• The proposed framework treats geo-localization as a clustering problem of local image features.
• Dominant Set clustering is utilized for efficient grouping of features, allowing dynamic selection of nearest neighbors.
• A second level of constrained Dominant Set clustering on global features ensures query image inclusion and bypasses heuristic voting.

Main Results:

• The method was evaluated on datasets of 102k and 300k street view images.
• It demonstrated superior performance, outperforming state-of-the-art methods by 20% and 7% on the respective datasets.
• The Dominant Set clustering approach proved orders of magnitude faster than existing techniques.

Conclusions:

• The novel geo-localization approach offers significant improvements in accuracy and computational efficiency.
• Dominant Set clustering provides a robust and scalable solution for large-scale image-based geo-localization.
• This framework advances the state-of-the-art in visual place recognition and autonomous navigation.