Scholarlink

<p>Cross-view geo-localization (CVGL) in real-world deployments suffers from orientation uncertainty: ground images are captured at arbitrary headings while aerial references remain North-aligned, causing severe retrieval instability under large azimuthal shifts. Existing approaches either rely on stochastic rotation augmentation, which does not explicitly enforce embedding consistency across viewing angles, or require explicit orientation labels and auxiliary prediction branches that increase model complexity. We propose DIRL (Direction-Inconsistency Robust Localization), a framework that models panoramic yaw rotation through cyclic panoramic shifts and implements rotation-consistent feature learning via multi-view contrastive training. DIRL introduces a symmetric Multi-View InfoNCE objective that jointly optimizes three complementary view pairs, combined with multi-rotation feature aggregation over uniformly sampled rotated variants to encourage consistent embeddings across heading changes, without requiring orientation labels or auxiliary prediction branches. Experiments on CVUSA, CVACT, VIGOR, and CV-Cities demonstrate that DIRL consistently improves rotation robustness: the mean Recall@1 degradation under arbitrary viewpoint rotations decreases from 11.10% to <b>3.98%</b> on CVUSA and from 11.64% to <b>3.46%</b> on CVACT, while competitive retrieval accuracy is maintained on both orientation-aligned and naturally misaligned benchmarks. These results demonstrate that enforcing rotation-consistent feature embeddings through multi-view contrastive learning provides an effective and lightweight solution for orientation-robust cross-view geo-localization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Geo-localization under viewpoint rotation: addressing direction inconsistencies with contrastive learning

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Youyuan Xue,
  • Kaiyi Lin,
  • Fan Guo,
  • Jin Tang,
  • Zhihu Wu

摘要

Cross-view geo-localization (CVGL) in real-world deployments suffers from orientation uncertainty: ground images are captured at arbitrary headings while aerial references remain North-aligned, causing severe retrieval instability under large azimuthal shifts. Existing approaches either rely on stochastic rotation augmentation, which does not explicitly enforce embedding consistency across viewing angles, or require explicit orientation labels and auxiliary prediction branches that increase model complexity. We propose DIRL (Direction-Inconsistency Robust Localization), a framework that models panoramic yaw rotation through cyclic panoramic shifts and implements rotation-consistent feature learning via multi-view contrastive training. DIRL introduces a symmetric Multi-View InfoNCE objective that jointly optimizes three complementary view pairs, combined with multi-rotation feature aggregation over uniformly sampled rotated variants to encourage consistent embeddings across heading changes, without requiring orientation labels or auxiliary prediction branches. Experiments on CVUSA, CVACT, VIGOR, and CV-Cities demonstrate that DIRL consistently improves rotation robustness: the mean Recall@1 degradation under arbitrary viewpoint rotations decreases from 11.10% to 3.98% on CVUSA and from 11.64% to 3.46% on CVACT, while competitive retrieval accuracy is maintained on both orientation-aligned and naturally misaligned benchmarks. These results demonstrate that enforcing rotation-consistent feature embeddings through multi-view contrastive learning provides an effective and lightweight solution for orientation-robust cross-view geo-localization.