Scholarlink

<p>Visual object tracking aims to continuously localize a target across video frames, yet remains challenging under complex background variations, occlusions, and deformations. Existing deep learning-based trackers often suffer from template contamination at initialization and passive historical modeling during tracking, leading to unstable localization. To address these issues, we propose FetaTrack, a foreground-aware tracking framework that reformulates dynamic template learning as a purification-driven and interaction-guided co-evolution process. Specifically, a foreground extraction (FE) module explicitly purifies the target representation at initialization, while a co-evolution mechanism within the encoder enables global interaction among the static template, dynamic template, and search region, allowing the template to be actively reshaped by context. Furthermore, a confidence-driven adaptive update strategy couples target state estimation with template evolution by selecting between mask-based correction and exponential moving average (EMA)-based adaptation, achieving a balance between robustness and flexibility. Extensive experiments on GOT-10k, TrackingNet, LaSOT, and LaSOText demonstrate that FetaTrack achieves competitive and consistent performance improvements under a variety of challenging tracking scenarios. Code and models are open-sourced at <a href="https://github.com/BestInYou/fetatrack">https://github.com/BestInYou/fetatrack</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

Fetatrack: Foreground-aware dynamic template co-evolution for visual object tracking

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Yibing Zhang,
  • Xuan Wang,
  • Yongchao Song,
  • Weiqing Yan,
  • Aoran Wang,
  • Zhe Dai

摘要

Visual object tracking aims to continuously localize a target across video frames, yet remains challenging under complex background variations, occlusions, and deformations. Existing deep learning-based trackers often suffer from template contamination at initialization and passive historical modeling during tracking, leading to unstable localization. To address these issues, we propose FetaTrack, a foreground-aware tracking framework that reformulates dynamic template learning as a purification-driven and interaction-guided co-evolution process. Specifically, a foreground extraction (FE) module explicitly purifies the target representation at initialization, while a co-evolution mechanism within the encoder enables global interaction among the static template, dynamic template, and search region, allowing the template to be actively reshaped by context. Furthermore, a confidence-driven adaptive update strategy couples target state estimation with template evolution by selecting between mask-based correction and exponential moving average (EMA)-based adaptation, achieving a balance between robustness and flexibility. Extensive experiments on GOT-10k, TrackingNet, LaSOT, and LaSOText demonstrate that FetaTrack achieves competitive and consistent performance improvements under a variety of challenging tracking scenarios. Code and models are open-sourced at https://github.com/BestInYou/fetatrack.