MobileGeo: Exploring Hierarchical Knowledge Distillation for Resource-Efficient Cross-view Drone Geo-Localization

TL;DR

MobileGeo tackles resource-efficient cross-view drone geo-localization (CVGL) for GNSS-denied settings by shifting complexity to training through Hierarchical Distillation (HD-CVGL)—combining Fine-Grained Inverse Self-Distillation, Uncertainty-Aware Prediction Alignment, and cross-distillation from a foundation model—and by introducing an inference-time Multi-view Selection Refinement Module (MSRM) that uses mutual information to select and fuse informative drone views. The training-time HD-CVGL yields a compact student without inference overhead, while MSRM reduces feature matching costs and boosts accuracy. Empirical results on University-1652 and SUES-200 show MobileGeo achieving state-of-the-art AP and Recall@1 with substantial FLOPs reductions (over 5×) and real-time edge-device performance (251.5 FPS on AGX Orin), including strong generalization in unsupervised domain adaptation and robustness under adverse weather and spatial offset. The work supplies a practical pathway for deploying accurate CVGL on mobile drones and edge devices, enabling reliable localization in GNSS-denied environments.

Abstract

Cross-view geo-localization (CVGL) enables drone localization by matching aerial images to geo-tagged satellite databases, which is critical for autonomous navigation in GNSS-denied environments. However, existing methods rely on resource-intensive feature alignment and multi-branch architectures, incurring high inference costs that limit their deployment on mobile edge devices. We propose MobileGeo, a mobile-friendly framework designed for efficient on-device CVGL. MobileGeo achieves its efficiency through two key components: 1) During training, a Hierarchical Distillation (HD-CVGL) paradigm, coupled with Uncertainty-Aware Prediction Alignment (UAPA), distills essential information into a compact model without incurring inference overhead. 2) During inference, an efficient Multi-view Selection Refinement Module (MSRM) leverages mutual information to filter redundant views and reduce computational load. Extensive experiments demonstrate that MobileGeo outperforms previous state-of-the-art methods, achieving a 4.19\% improvement in AP on University-1652 dataset while being over 5$\times$ more efficient in FLOPs and 3$\times$ faster. Crucially, MobileGeo runs at 251.5 FPS on an NVIDIA AGX Orin edge device, demonstrating its practical viability for real-time on-device drone geo-localization.

Figures (14)

• Figure 1: Dual-perspective efficiency analysis of our MobileGeo on University-1652 Drone$\rightarrow$Satellite benchmark. (a) Runtime efficiency: Our method achieves 97.15% R@1 with 1022 FPS, enabling real-time mobile applications. (b) Computational efficiency: With only 4.45G FLOPs, our approach surpasses heavier models ($>$20G FLOPs) in accuracy. Our method consistently dominates existing approaches in both computational and runtime efficiency while achieving state-of-the-art performance. $*$ denotes the efficient model after hierarchical distillation, $^\dagger$ indicates the model with post-process.
• Figure 2: The top panel illustrates the workflow for cross-view drone geo-localization in a GPS-signal-denied environment. The bottom panel contrasts existing "mobile-unfriendly" approaches with our proposed "mobile-friendly" MobileGeo method. (a) Illustration of prior methods chen2025multichen2024sdplxia2024enhancingdu2024ccr that introduce auxiliary modules during feature extraction. (b) Our proposed module is inference-free, incurring no additional computational overhead at deployment. Furthermore, in the feature matching stage, our MSRM significantly reduces computational complexity by selectively filtering and fusing multi-view features.
• Figure 3: Overview of our MobileGeo framework. (a) The Hierarchical Distillation for CVGL (HD-CVGL). This is a two-step process: first, a tiny student model undergoes Fine-Grained Inverse Self-distillation. Second, the fine-tuned foundation model acts as a teacher, providing guidance at both the feature and logit levels. (b) The inference stage pipeline. The Multi-view Selection Refinement Module (MSRM) leverages mutual information to select discriminative drone images from multiple views, effectively boosting both retrieval accuracy and speed. (c) A detailed illustration of the self-distillation pipeline, depicting the flow of optimization objectives. It incorporates an Uncertainty-Aware Prediction Alignment (UAPA) mechanism to mitigate challenges from data imbalance.
• Figure 4: Analysis of cross-view performance dynamics. The figure illustrates the training dynamics of our model, plotting the accuracy for satellite (blue) and drone (orange) domains against training epochs. A noticeable performance discrepancy, or 'Cross-view Gap' (shaded region), emerges where the satellite view consistently outperforms the drone view. Critically, our analysis reveals that this gap is not static but exhibits a clear widening trend during training in the first several epochs, highlighted by the purple dashed line which plots the gap's magnitude during its periods of increase.
• Figure 5: Overview of our Multi-view Selection Refinement Module (MSRM). On the left, we showcase the feature matching process between multi-view drone images (e.g., from the gallery database) and a satellite image. On the right, the detailed pipeline of MSRM is presented. The process begins with the construction of a multi-view drone descriptor database. As shown, features captured from the same viewpoint are modeled as a Gaussian distribution. The variables $h=[h_1,h_2,h_3]$ and $\theta$ represent the drone's spatial position. Each selected sample is a 768-dimensional vector. The operators $\otimes$, $\odot$, and $\oplus$ denote element-wise multiplication, dot product, and addition, respectively. $E$ represents the Euclidean distance, while $Q_*$$/$$G_*$ denote a query $/$ gallery sample.