Visual Reasoning for Ground to Aerial Localization

This paper introduces ViReLoc, a novel framework for ground-to-aerial localization using only visual representations. It addresses the limitations of text-based reasoning in spatial tasks by learning spatial dependencies and geometric relations directly from visual data. The use of reinforcement learning and contrastive learning for cross-view alignment is a key aspect. The work's significance lies in its potential for secure navigation solutions without relying on GPS data.

• •Proposes ViReLoc, a visual reasoning framework for ground-to-aerial localization.
• •Utilizes visual representations for planning and localization, avoiding reliance on text-based reasoning.
• •Employs reinforcement learning and contrastive learning for improved spatial reasoning and cross-view alignment.
• •Demonstrates potential for secure navigation without GPS.