← 研踪 PaperTrailarXiv:2611.01074 · v2
Grounding Across Views: Cross-View Referring Reasoning for Visual Grounding
首次提交 2026-11-01 00:12
版本更新 2026-11-17 21:04
雷达观测 2026-11-18 19:41
摘要
Cross-view object geo-localization (CVOGL) aims to identify a query-view target in a satellite image, yet reliable instance grounding remains challenging under substantial viewpoint discrepancies. Although recent methods incorporate multimodal prompts, they insufficiently address the varying transferability of linguistic and visual evidence across views. Viewpoint-dependent descriptions (e.g., building facades visible in oblique UAV views but not in top-down satellite images) may become unreliable for satellite grounding, while visually similar instances further complicate target identification. To address these challenges, we rethink CVOGL from a visual grounding perspective and propose GAV (Grounding Across Views), a view-aware framework for constructing target-centered evidence and establishing discriminative cross-view correspondence. Specifically, view-aware referring reasoning mitigates viewpoint-dependent ambiguity, multi-token exemplar prompting and contrastive alignment strengthen visual-semantic correspondence, and scene-level supervision with alignment-based verification distinguishes competing satellite instances. We further introduce CVOGL-Text, a street- and drone-to-satellite benchmark with query-derived expressions. GAV achieves state-of-the-art performance on MoP-UAV with 62.55% Acc@0.5 (+18.19 percentage points) and outperforms evaluated click-point baselines on CVOGL-Text.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 21:04 · 最新观测版本
Grounding Across Views: Cross-View Referring Reasoning for Visual Grounding
Cross-view object geo-localization (CVOGL) aims to identify a query-view target in a satellite image, yet reliable instance grounding remains challenging under substantial viewpoint discrepancies. Although recent methods incorporate multimodal prompts, they insufficiently address the varying transferability of linguistic and visual evidence across views. Viewpoint-dependent descriptions (e.g., building facades visible in oblique UAV views but not in top-down satellite images) may become unreliable for satellite grounding, while visually similar instances further complicate target identification. To address these challenges, we rethink CVOGL from a visual grounding perspective and propose GAV (Grounding Across Views), a view-aware framework for constructing target-centered evidence and establishing discriminative cross-view correspondence. Specifically, view-aware referring reasoning mitigates viewpoint-dependent ambiguity, multi-token exemplar prompting and contrastive alignment strengthen visual-semantic correspondence, and scene-level supervision with alignment-based verification distinguishes competing satellite instances. We further introduce CVOGL-Text, a street- and drone-to-satellite benchmark with query-derived expressions. GAV achieves state-of-the-art performance on MoP-UAV with 62.55% Acc@0.5 (+18.19 percentage points) and outperforms evaluated click-point baselines on CVOGL-Text.
查看此版本解读 →查看此版本原文 ↗