Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Dual-level Progressive Hardness-Aware Reweighting for Cross-View Geo-Localization

Guozheng Zheng, Jian Guan, Mingjie Xie, Xuanjia Zhao, Congyi Fan, Shiheng Zhang, Pengming Feng

TL;DR

CVGL between drone and satellite imagery suffers from severe viewpoint gaps and hard negatives, and static weighting can yield unstable gradients. The paper proposes Dual-level Progressive Hardness-aware Reweighting (DPHR) consisting of a Ratio-Based Difficulty-Aware (RDA) module and a Progressive Adaptive Loss Weighting (PALW) mechanism. RDA computes the relative hardness score $h_{i,k}=d(p_i,q_i)/(d(p_i,q_i)+d(q_i,n_{i,k}))$ and maps it to a weight $\omega_{i,k}$ to emphasize harder negatives, while PALW uses a training-progress signal to adapt the loss via $\mathcal{L}_{DPHR}=\mathcal{L}_{tri}+\lambda^{t}\mathcal{L}_{wtri}$ with $\lambda^{t}$ updated from $\hat{\alpha}_t$. Experiments on University-1652 and SUES-200 show consistent improvements in Recall@1 and Average Precision across retrieval directions and drone altitudes, and ablation demonstrates the two components are complementary.

Abstract

Cross-view geo-localization (CVGL) between drone and satellite imagery remains challenging due to severe viewpoint gaps and the presence of hard negatives, which are visually similar but geographically mismatched samples. Existing mining or reweighting strategies often use static weighting, which is sensitive to distribution shifts and prone to overemphasizing difficult samples too early, leading to noisy gradients and unstable convergence. In this paper, we present a Dual-level Progressive Hardness-aware Reweighting (DPHR) strategy. At the sample level, a Ratio-based Difficulty-Aware (RDA) module evaluates relative difficulty and assigns fine-grained weights to negatives. At the batch level, a Progressive Adaptive Loss Weighting (PALW) mechanism exploits a training-progress signal to attenuate noisy gradients during early optimization and progressively enhance hard-negative mining as training matures. Experiments on the University-1652 and SUES-200 benchmarks demonstrate the effectiveness and robustness of the proposed DPHR, achieving consistent improvements over state-of-the-art methods.

Dual-level Progressive Hardness-Aware Reweighting for Cross-View Geo-Localization

TL;DR

CVGL between drone and satellite imagery suffers from severe viewpoint gaps and hard negatives, and static weighting can yield unstable gradients. The paper proposes Dual-level Progressive Hardness-aware Reweighting (DPHR) consisting of a Ratio-Based Difficulty-Aware (RDA) module and a Progressive Adaptive Loss Weighting (PALW) mechanism. RDA computes the relative hardness score and maps it to a weight to emphasize harder negatives, while PALW uses a training-progress signal to adapt the loss via with updated from . Experiments on University-1652 and SUES-200 show consistent improvements in Recall@1 and Average Precision across retrieval directions and drone altitudes, and ablation demonstrates the two components are complementary.

Abstract

Cross-view geo-localization (CVGL) between drone and satellite imagery remains challenging due to severe viewpoint gaps and the presence of hard negatives, which are visually similar but geographically mismatched samples. Existing mining or reweighting strategies often use static weighting, which is sensitive to distribution shifts and prone to overemphasizing difficult samples too early, leading to noisy gradients and unstable convergence. In this paper, we present a Dual-level Progressive Hardness-aware Reweighting (DPHR) strategy. At the sample level, a Ratio-based Difficulty-Aware (RDA) module evaluates relative difficulty and assigns fine-grained weights to negatives. At the batch level, a Progressive Adaptive Loss Weighting (PALW) mechanism exploits a training-progress signal to attenuate noisy gradients during early optimization and progressively enhance hard-negative mining as training matures. Experiments on the University-1652 and SUES-200 benchmarks demonstrate the effectiveness and robustness of the proposed DPHR, achieving consistent improvements over state-of-the-art methods.

Paper Structure

This paper contains 11 sections, 10 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Proposed Method
  3. Preliminaries
  4. Ratio-Based Difficulty-Aware Module
  5. Progressive Adaptive Loss Weighting Mechanism
  6. Experiments
  7. Experimental Setup
  8. Performance Comparison
  9. Ablation Study
  10. Visualization Analysis
  11. Conclusion

Figures (3)

  • Figure 1: Illustration of visual similarity in feature space for CVGL. Given a query image, the challenge arises when the hard negative, due to its structural and color similarities, becomes closer to the query in feature space than the true positive, misguiding the model into making incorrect matches.
  • Figure 2: The overall framework of the proposed DPHR strategy for CVGL, which consists of two key components, i.e., ratio-aware difficulty-aware (RDA) module and progressive adaptive loss weighting (PALW) mechanism. Here, RDA module assigns sample-level weights based on relative hardness, while the PALW mechanism adaptively regulates the overall loss contribution according to training progress.
  • Figure 3: The t-SNE visualization comparing MCCG and MCCG-DPHR. Our strategy improves separation between queries and hard negatives, demonstrating its effectiveness in handling challenging negative samples.