Scholarlink

<p>In aquaculture, precise monitoring of fish behavior is crucial for breeding superior strains. Traditional tracking methods face challenges such as occlusion, deformation, and complex underwater environments. To address these, we propose DiffFish, a diffusion model-based framework that unifies object detection and association for multi-object fish tracking. DiffFish employs dynamic fish mask insertion and a random frame-skipping strategy to enhance robustness and reduce trajectory fragmentation. Experimental results on the LC-MOT dataset demonstrate DiffFish’s superior performance with 93.9% MOTA and 75.6% IDF1, outperforming existing methods. This approach provides an efficient solution for analyzing fish motion and behavior in aquaculture. The dataset and codes are publicly available at https://github.com/arya7bling-blip/DiffFish.</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

DiffFish: a unified diffusion model for robust multi-object fish tracking in aquaculture

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Kexin Yuan,
  • Yunchen Tian,
  • Jianing Quan

摘要

In aquaculture, precise monitoring of fish behavior is crucial for breeding superior strains. Traditional tracking methods face challenges such as occlusion, deformation, and complex underwater environments. To address these, we propose DiffFish, a diffusion model-based framework that unifies object detection and association for multi-object fish tracking. DiffFish employs dynamic fish mask insertion and a random frame-skipping strategy to enhance robustness and reduce trajectory fragmentation. Experimental results on the LC-MOT dataset demonstrate DiffFish’s superior performance with 93.9% MOTA and 75.6% IDF1, outperforming existing methods. This approach provides an efficient solution for analyzing fish motion and behavior in aquaculture. The dataset and codes are publicly available at https://github.com/arya7bling-blip/DiffFish.