Robust Drone-View Geo-Localization via Content-Viewpoint Disentanglement

TL;DR

This work tackles drone-view geo-localization (DVGL) by addressing viewpoint-induced conflicts in cross-view matching. It introduces CVD, a Content-Viewpoint Disentanglement framework that models the cross-view feature space as a composite manifold $\mathcal{M}$ decomposed into content $\mathcal{M}_c$ and viewpoint $\mathcal{M}_v$ submanifolds, and employs an embed-disentangle-reconstruct pipeline. CVD enforces two constraints: (i) an intra-view independence constraint that minimizes mutual information between content and viewpoint via the Sliced Wasserstein Distance, and (ii) an inter-view reconstruction constraint that swaps content and viewpoint across paired views to preserve factor-specific semantics, complemented by a standard InfoNCE loss for content alignment. Extensive experiments on University-1652, SUES-200, CVUSA, and CVACT demonstrate consistent improvements across backbones, enhanced robustness to viewpoint and altitude variation, and improved cross-dataset generalization, all while reducing inference latency due to factorized representations. Overall, explicit content-viewpoint disentanglement yields more robust and scalable DVGL systems with practical deployment benefits.

Abstract

Drone-view geo-localization (DVGL) aims to match images of the same geographic location captured from drone and satellite perspectives. Despite recent advances, DVGL remains challenging due to significant appearance changes and spatial distortions caused by viewpoint variations. Existing methods typically assume that drone and satellite images can be directly aligned in a shared feature space via contrastive learning. Nonetheless, this assumption overlooks the inherent conflicts induced by viewpoint discrepancies, resulting in extracted features containing inconsistent information that hinders precise localization. In this study, we take a manifold learning perspective and model $\textit{the feature space of cross-view images as a composite manifold jointly governed by content and viewpoint}$. Building upon this insight, we propose $\textbf{CVD}$, a new DVGL framework that explicitly disentangles $\textit{content}$ and $\textit{viewpoint}$ factors. To promote effective disentanglement, we introduce two constraints: $\textit{(i)}$ an intra-view independence constraint that encourages statistical independence between the two factors by minimizing their mutual information; and $\textit{(ii)}$ an inter-view reconstruction constraint that reconstructs each view by cross-combining $\textit{content}$ and $\textit{viewpoint}$ from paired images, ensuring factor-specific semantics are preserved. As a plug-and-play module, CVD integrates seamlessly into existing DVGL pipelines and reduces inference latency. Extensive experiments on University-1652 and SUES-200 show that CVD exhibits strong robustness and generalization across various scenarios, viewpoints and altitudes, with further evaluations on CVUSA and CVACT confirming consistent improvements.

Figures (3)

• Figure 1: Comparison between previous methods and our CVD. Left: Existing methods can be interpreted as operating on a single manifold $\mathcal{M}$, where contrastive objectives directly pull positive pairs closer and push negative pairs away. Right: Our method learns disentangled representations by mapping inputs onto two submanifolds corresponding to content$\mathcal{M}_c$ and viewpoint$\mathcal{M}_v$. This separation is enforced via two constraints (see \ref{['sec:independence']} and \ref{['sec:reconstruction']}), promoting effective disentanglement and thereby enhancing cross-view matching performance.
• Figure 2: Overview of the proposed CVD.
• Figure 3: Qualitative results of cross-view reconstruction on the University-1652 and SUES-200 datasets.