计算机科学
人工智能
人机交互
可视化
自然语言处理
数据可视化
特征(语言学)
计算机视觉
语义学(计算机科学)
语音识别
作者
Mingxin Yu,Xiang Zhou,Yiyuan Ge,Mingwei Lin,Xin Luo
标识
DOI:10.1109/tmm.2026.3734743
摘要
Aerial-Ground Person Re-identification aims to match individuals across drastically different viewpoints, posing challenges due to significant variations in appearance, scale, and background context. Existing visual-based methods struggle to bridge the cross-view domain gap and often overlook semantic cues that could facilitate alignment. To address these limitations, we propose Mapping View-bias with Prompts (MVP), a two-stage vision-language framework that models and alleviates viewpoint bias for more robust aerial-ground representation learning. Specifically, MVP first models viewpoint bias through view-aware prompt learning and then employs view-agnostic textual guidance to learn more discriminative identity representations. To further enhance cross-view alignment, we introduce the Aerial-Ground Similarity Consistency loss, which preserves relational consistency between aerial and ground representations across visual and textual modalities. This progressive design enables the model to exploit semantic guidance while reducing reliance on viewpoint-specific cues, resulting in identity features that are more robust to severe viewpoint variations. Extensive experiments on the CARGO, AG-ReID, AG-ReID.v2, and LAGPeR datasets demonstrate that MVP achieves best or near-best performance across multiple synthetic and real-world AGPReID protocols, with clear gains on CARGO, AG-ReID.v2, and LAGPeR, and maintains strong cross-dataset generalization under real-world aerial-ground transfer settings.
科研通智能强力驱动
Strongly Powered by AbleSci AI