Fugu-MT: arxivの論文翻訳

Towards Active Cross-View Object Geo-Localization

Authors Shunyu Yao, Xiaohan Zhang, Zhuoran Yang, Haoqi Lai, Qi Ming, Xiaoxi Hu, Hui-Liang Shen, Si-Yuan Cao
Affiliations Zhejiang University / Beijing University of Technology / Tsinghua University
Categories Method / Active Learning / Sequential view selection and stopping criteria, Application / Object Localization / Cross-view geo-localization in UAV imagery, Evaluation / Model Evaluation / Performance comparison using minimal query views
License CC BY 4.0

Abstract Overview

This paper introduces ActiveGeo, a reformulation of cross-view object geo-localization in which a mobile agent sequentially acquires additional viewpoints and decides when to stop, rather than relying on a single fixed query image. To address this setting, the authors propose ActiveMoPT, built on MoPT with a three-stage pipeline: multi-view prompt-preserving adaptation, trajectory-guided policy initialization, and cost-aware policy refinement. The method is designed to preserve the initial target prompt across later unprompted views while jointly learning viewpoint selection and stopping behavior. The authors also construct ActiveGeo-858, a zero-shot benchmark with 858 scenes and 1,716 target annotations to evaluate generalization beyond MoP-UAV.

Novelty

The primary novelty is extending passive cross-view object geo-localization to an active decision-making framework where the agent dynamically selects informative viewpoints and learns when to stop. Methodologically, it introduces prompt-preserving multi-view adaptation alongside a cost-aware policy trained via trajectory initialization and reinforcement learning refinement.

Results

ActiveMoPT outperforms passive baselines with minimal query views, achieving 0.384 mIoU on MoP-UAV using 1.291 average views under point prompts compared to 0.368 for MoPT with one view. On zero-shot ActiveGeo-858, it reaches 0.391 mIoU with 1.195 average views versus 0.381 for MoPT. Policy refinement via GRPO further reduces average views on MoP-UAV from 1.886 to 1.450 while improving localization to 0.414 mIoU under multimodal prompts.

Key Points

  1. ActiveGeo reformulates cross-view object geo-localization as an active sequential decision task with dynamic viewpoint selection and stopping criteria.
  2. ActiveMoPT uses a three-stage framework that aggregates subsequent unprompted views while preserving the initial prompt identity.
  3. The model achieves state-of-the-art accuracy with low view budgets on MoP-UAV and demonstrates zero-shot transfer on the new ActiveGeo-858 dataset.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.