Scholarlink

<p>Open-vocabulary detection (OVD) and open-world object detection (OWOD) are increasingly important for unmanned aerial vehicle (UAV) perception, where detectors must recognize known, unseen, long-tail, and unknown objects in complex aerial scenes. UAV-based detection is particularly challenging due to small targets, cluttered backgrounds, large viewpoint changes, and strict efficiency requirements for onboard deployment. Existing convolutional detectors capture local details efficiently but are limited in long-range dependency modeling, while Transformer-based methods provide stronger global reasoning at the cost of high computational complexity. In addition, many current OVD and OWOD methods rely on shallow cross-modal fusion, which weakens semantic consistency between aerial visual regions and textual category descriptions. To address these challenges, we propose MamWorld, a unified state-space-based detection framework for UAV-oriented open-category object detection. MamWorld introduces an ODMamba-based backbone with MassBlock to enhance cross-channel interaction, multi-scale spatial aggregation, and long-range dependency modeling for small and ambiguous aerial objects. It further employs TextMambaBlock and SGSS-TextMambaBlock to generate semantically selective textual representations, and MambaFusion-PAN to perform recursive bidirectional vision–language fusion across scales. This end-to-end design enables efficient alignment between UAV imagery and open-category semantics while supporting both OVD and OWOD. Experiments on LVIS, M-OWODB, and S-OWODB demonstrate that MamWorld achieves strong performance, especially for rare and unknown categories, while maintaining practical efficiency for UAV perception scenarios.</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

MamWorld for open-world object detection with state space modeling and cross-modal fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月11日

查看全文
  • Zongqiang Deng,
  • Hongfei Zhao

摘要

Open-vocabulary detection (OVD) and open-world object detection (OWOD) are increasingly important for unmanned aerial vehicle (UAV) perception, where detectors must recognize known, unseen, long-tail, and unknown objects in complex aerial scenes. UAV-based detection is particularly challenging due to small targets, cluttered backgrounds, large viewpoint changes, and strict efficiency requirements for onboard deployment. Existing convolutional detectors capture local details efficiently but are limited in long-range dependency modeling, while Transformer-based methods provide stronger global reasoning at the cost of high computational complexity. In addition, many current OVD and OWOD methods rely on shallow cross-modal fusion, which weakens semantic consistency between aerial visual regions and textual category descriptions. To address these challenges, we propose MamWorld, a unified state-space-based detection framework for UAV-oriented open-category object detection. MamWorld introduces an ODMamba-based backbone with MassBlock to enhance cross-channel interaction, multi-scale spatial aggregation, and long-range dependency modeling for small and ambiguous aerial objects. It further employs TextMambaBlock and SGSS-TextMambaBlock to generate semantically selective textual representations, and MambaFusion-PAN to perform recursive bidirectional vision–language fusion across scales. This end-to-end design enables efficient alignment between UAV imagery and open-category semantics while supporting both OVD and OWOD. Experiments on LVIS, M-OWODB, and S-OWODB demonstrate that MamWorld achieves strong performance, especially for rare and unknown categories, while maintaining practical efficiency for UAV perception scenarios.