Work

BiCrossNet: resource-efficient cross-view geolocalization with binary neural networks

Year: 2025

Type: article

Abstract: This paper presents BiCrossNet, a novel approach to cross-view geolocalization utilizing binary neural networks to significantly reduce computational complexity while maintaining competitive performance. Key contributions include the development of a Bi-Gradual Unfreezing method to enhance transfer learning, a Bi-Partitioned Optimization strategy to improve training stability, and the use of logit-based knowledge distillation to supplement standard losses. Experimental results on the University-1652 and SUES-200 datasets demonstrate that BiCrossNet establishes a new benchmark in the efficiency-performance trade-off. It achieves up to a 90.87-fold reduction in operations and uses 4.64 times less disk space compared to similar-performing state-of-the-art models on the SUES-200 dataset, and a 30-fold reduction in operations and 5.13 times less disk space on the University-1652 dataset. The code is available at https://anonymous.4open.science/r/BiCrossNet-FB7A/README.md .

Source: Machine Learning Science and Technology

DOI: 10.1088/2632-2153/adfa67

Authors: Federico Fontana, Thomas Jantos, Jan Steinbrener, Luigi Cinque, Gian Luca Foresti, Bernhard Rinner

Institutions: Sapienza University of Rome, University of Klagenfurt, University of Udine

Language: English

Topic: Advanced Image and Video Retrieval Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $3,000