Scholarlink

<p>Comprehensive spatiotemporal context is essential for robust visual object tracking. However, existing methods often restrict the flow of spatiotemporal information to fixed-length or single-stage representations, which limits the amount of information passed across time and constrains the ability to capture target variations. To address this limitation, we propose MambaCPT, a video-level tracker that fully exploits multi-stage spatiotemporal content through dual prompting. Unlike prior methods that update only explicit prompts, MambaCPT further introduces implicit prompts to complement explicit ones at each stage. Specifically, explicit and implicit prompts are updated together with current search features through the proposed Mamba Prompt module, enabling the model to effectively capture the dynamic variations of the target. The updated key information is stored in the hidden states of Mamba to construct a temporally coherent contextual representation. This spatiotemporal context is further integrated into the attention mechanism to enhance the feature modeling of both the template and the search. By leveraging the sequential modeling capability of Mamba and the strong appearance modeling of attention, MambaCPT effectively captures and utilizes multi-level spatiotemporal dependencies. Compared with recent trackers, our MambaCPT achieves superior performance across seven challenging benchmarks, particularly on GOT-10k, while maintaining real-time speed. The source code and models are released on <a href="https://github.com/xiaomengxin123/MambaCPT">https://github.com/xiaomengxin123/MambaCPT</a>.</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

Mamba-driven context-aware tracking with dual prompts

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zhao Huang,
  • Lei Liu,
  • Shuai Wang,
  • Jun Wang

摘要

Comprehensive spatiotemporal context is essential for robust visual object tracking. However, existing methods often restrict the flow of spatiotemporal information to fixed-length or single-stage representations, which limits the amount of information passed across time and constrains the ability to capture target variations. To address this limitation, we propose MambaCPT, a video-level tracker that fully exploits multi-stage spatiotemporal content through dual prompting. Unlike prior methods that update only explicit prompts, MambaCPT further introduces implicit prompts to complement explicit ones at each stage. Specifically, explicit and implicit prompts are updated together with current search features through the proposed Mamba Prompt module, enabling the model to effectively capture the dynamic variations of the target. The updated key information is stored in the hidden states of Mamba to construct a temporally coherent contextual representation. This spatiotemporal context is further integrated into the attention mechanism to enhance the feature modeling of both the template and the search. By leveraging the sequential modeling capability of Mamba and the strong appearance modeling of attention, MambaCPT effectively captures and utilizes multi-level spatiotemporal dependencies. Compared with recent trackers, our MambaCPT achieves superior performance across seven challenging benchmarks, particularly on GOT-10k, while maintaining real-time speed. The source code and models are released on https://github.com/xiaomengxin123/MambaCPT.