Scholarlink

<p>Real-time environmental monitoring of inland waters remains a challenge due to the lack of domain-specific datasets for instance segmentation. This paper presents SWIRIS, a dataset consisting of 3011 high-resolution images curated for unmanned surface vehicle (USV) applications for monitoring water quality. SWIRIS targets five critical environmental classes: <i>Algae</i>, <i>Water</i>, <i>Water Edge</i>, <i>Reeds</i>, and <i>Background Vegetation</i>. Beyond providing a labeled benchmark, we explore the efficacy of generative AI for dataset expansion; by applying Stable Diffusion and ControlNet for synthetic data augmentation, we observed significant performance gains, with AP50:95 increasing by up to 40% for YOLO-based models. To bridge the gap between training and deployment, we conduct a comprehensive power-performance analysis on the NVIDIA Jetson Orin Nano platform. Our results demonstrate that real-time constraints (15 FPS) can be maintained even in “extreme low-power” modes, yielding between 56 and 76% energy savings crucial to real autonomous missions. SWIRIS serves as a foundational resource for developing robust, energy-aware perception systems in continental aquatic monitoring.</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

SWIRIS: a stereo vision dataset for instance segmentation in inland water reservoirs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月19日

查看全文
  • Jose Luis Mela,
  • Carlos García Sánchez

摘要

Real-time environmental monitoring of inland waters remains a challenge due to the lack of domain-specific datasets for instance segmentation. This paper presents SWIRIS, a dataset consisting of 3011 high-resolution images curated for unmanned surface vehicle (USV) applications for monitoring water quality. SWIRIS targets five critical environmental classes: Algae, Water, Water Edge, Reeds, and Background Vegetation. Beyond providing a labeled benchmark, we explore the efficacy of generative AI for dataset expansion; by applying Stable Diffusion and ControlNet for synthetic data augmentation, we observed significant performance gains, with AP50:95 increasing by up to 40% for YOLO-based models. To bridge the gap between training and deployment, we conduct a comprehensive power-performance analysis on the NVIDIA Jetson Orin Nano platform. Our results demonstrate that real-time constraints (15 FPS) can be maintained even in “extreme low-power” modes, yielding between 56 and 76% energy savings crucial to real autonomous missions. SWIRIS serves as a foundational resource for developing robust, energy-aware perception systems in continental aquatic monitoring.