Scholarlink

<p>The past decades have witnessed significant progress in visual object tracking, and many advanced trackers have been proposed. However, previous research on object tracking was generally conducted in general-purpose scenarios, with few specifically targeting wild animal tracking (WAT), which seriously hinders the progress of computer vision research in wild animals. Therefore, how to effectively leverage the power of the existing advanced trackers to solve the challenges in WAT becomes a challenging problem. Furthermore, the lack of high-quality wildlife training data is a major obstacle to advancing wild animal tracking, as training an effective tracking model needs the support of large-scale and high-quality training data. To address these issues, this paper introduces a new Wild Animal Tracking framework based on the high-quality Segment anything model (SAM) and Domain Adaptation (WATS-DA), which aims to provide new motivation and direction for wild animal-oriented object tracking research. Specifically, we provide a high-quality SAM (HQ-SAM)-based data processing method to generate many high-quality target domain training samples from raw wild animal images. A domain adaptation-based training framework is introduced to migrate existing trackers’ power to WAT. To solve the long-term WAT problem, we devise a plug-and-play redetection module to recover tracking from failure. Moreover, to mitigate the data scarcity problem in WAT, we build a large-scale dataset named Wildlife2024 for the training and testing of domain adaptation trackers in WAT. Extensive experimental results on WAT benchmark datasets attest to the effectiveness of the proposed tracking framework. Codes and datasets are available on <a href="https://github.com/Hgg12/WATS-DA">https://github.com/Hgg12/WATS-DA</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

Wild Animal Tracking with High-Quality Segment Anything Model and Domain Adaptation

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Ganggang Huang,
  • Fasheng Wang,
  • Binbin Wang,
  • Hanwei Li,
  • Mingshu Zhang,
  • Mengyin Wang,
  • Fuming Sun,
  • Haojie Li

摘要

The past decades have witnessed significant progress in visual object tracking, and many advanced trackers have been proposed. However, previous research on object tracking was generally conducted in general-purpose scenarios, with few specifically targeting wild animal tracking (WAT), which seriously hinders the progress of computer vision research in wild animals. Therefore, how to effectively leverage the power of the existing advanced trackers to solve the challenges in WAT becomes a challenging problem. Furthermore, the lack of high-quality wildlife training data is a major obstacle to advancing wild animal tracking, as training an effective tracking model needs the support of large-scale and high-quality training data. To address these issues, this paper introduces a new Wild Animal Tracking framework based on the high-quality Segment anything model (SAM) and Domain Adaptation (WATS-DA), which aims to provide new motivation and direction for wild animal-oriented object tracking research. Specifically, we provide a high-quality SAM (HQ-SAM)-based data processing method to generate many high-quality target domain training samples from raw wild animal images. A domain adaptation-based training framework is introduced to migrate existing trackers’ power to WAT. To solve the long-term WAT problem, we devise a plug-and-play redetection module to recover tracking from failure. Moreover, to mitigate the data scarcity problem in WAT, we build a large-scale dataset named Wildlife2024 for the training and testing of domain adaptation trackers in WAT. Extensive experimental results on WAT benchmark datasets attest to the effectiveness of the proposed tracking framework. Codes and datasets are available on https://github.com/Hgg12/WATS-DA.