Scholarlink

<p>Markerless pose estimation has emerged as a promising methodology for quantifying the behavior of freely moving mice. However, achieving scientifically precise 2D Mouse Pose Estimation (MPE) remains challenging, primarily due to the scarcity of large scale benchmark datasets and the underdevelopment of techniques tailored to animal behavior experiments. Existing pose estimation techniques developed for Human Pose Estimation (HPE) are rarely directly transferable to mice. One key distinction between HPE and MPE stems from the stricter precision requirements in animal behavior studies relative to human centric scenarios. In this work, we propose a novel framework that integrates Multi-Scale Context (MSC) with Sensitive-Aware Loss (SAL) to tackle this challenge. Specifically, the MSC exploits multi-scale contextual information to capture discriminative keypoint representations, while the SAL alleviates the extreme class imbalance inherently encountered in precise keypoint localization, thereby facilitating accurate localization. Experiments conducted on two public real world datasets demonstrate that our approach achieves accurate mouse keypoint localization under extremely low illumination conditions (0–15 lux).</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

Precise 2D mouse pose estimation via multi-scale context and sensitive-aware loss from low illumination environment

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Yubin Geng,
  • Jiaxin Deng,
  • Zhicheng Wang,
  • Junbiao Pang

摘要

Markerless pose estimation has emerged as a promising methodology for quantifying the behavior of freely moving mice. However, achieving scientifically precise 2D Mouse Pose Estimation (MPE) remains challenging, primarily due to the scarcity of large scale benchmark datasets and the underdevelopment of techniques tailored to animal behavior experiments. Existing pose estimation techniques developed for Human Pose Estimation (HPE) are rarely directly transferable to mice. One key distinction between HPE and MPE stems from the stricter precision requirements in animal behavior studies relative to human centric scenarios. In this work, we propose a novel framework that integrates Multi-Scale Context (MSC) with Sensitive-Aware Loss (SAL) to tackle this challenge. Specifically, the MSC exploits multi-scale contextual information to capture discriminative keypoint representations, while the SAL alleviates the extreme class imbalance inherently encountered in precise keypoint localization, thereby facilitating accurate localization. Experiments conducted on two public real world datasets demonstrate that our approach achieves accurate mouse keypoint localization under extremely low illumination conditions (0–15 lux).